CoMap: Mapping Contagion in the Euro

Area Banking Sector

Giovanni Covi*, Mehmet Ziya Gorpe**, Christoffer Kok***

Disclaimer: This Paper should not be reported as representing the views of the European

Central Bank (ECB) and of the International Monetary Fund (IMF). The views expressed

are those of the authors and do not necessarily reflect those of the ECB and the IMF.

Acknowledgments: We wish to thank Cédric Tille, Henry Jérôme, Daniel Hardy, Mattia

Montagna, Tuomas Peltonen, Alberto Giovannini, Barbara Meller, Eric Schaanning,

Andrei Sarechev, Niki Anderson, Caterina Lepore, Nuno Silva, Noam Michelson, and

participants of the ASSA 2019 Meeting, the Systemic Risk and Macroprudential CEPR

Conference, the Complex Networks 2018 Conference, the MBF 2018 Conference Rome,

6th

ECB/CBRT Conference, Credit 2018 Conference Venice, the Banco de

Portugal/ECB/ESRB Workshop, the ESRB Interconnectedness Workgroup, Bank of

England Fire-sales Workshop and ECB and IMF internal seminars for helpful comments

and suggestions.

* European Central Bank, Macroprudential Policy and Financial Stability Directorate,

Stress Test Modelling Division, Frankfurt am Main. Email: giovanni.covi@ecb.europa.eu

** International Monetary Fund, Monetary and Capital Markets Department, Washington

DC. Email: mgorpe@imf.org

*** European Central Bank, Macroprudential Policy and Financial Stability Directorate,

Stress Test Modelling Division, Frankfurt am Main. Email: christoffer.kok@ecb.int

2

Abstract

This paper presents a novel approach to investigate and model the network of euro

area banks’ large exposures within the global banking system. Drawing on a unique

dataset, the paper documents the degree of interconnectedness and systemic risk of

the euro area banking system based on bilateral linkages. We then develop a

Contagion Mapping (CoMap) methodology to study contagion potential of an

exogenous default shock via counterparty credit and funding risks. We construct

contagion and vulnerability indices measuring respectively the systemic importance

of banks and their degree of fragility. Decomposing the results into the respective

contributions of credit and funding shocks provides insights to the nature of

contagion which can be used to calibrate bank-specific capital and liquidity

requirements and large exposures limits. We find that tipping points shifting the euro

area banking system from a less vulnerable state to a highly vulnerable state are a

non-linear function of the combination of network structures and bank-specific

characteristics.

Keywords: Systemic Risk, Network Analysis, Interconnectedness, Large Exposures,

Stress Test, Macroprudential Policy.

JEL Codes: D85, G17, G33, L14.

3

Non-technical Summary

We develop a contagion mapping methodology (CoMap) to study systemic risk

stemming from interconnectedness based on the euro area Significant Institutions’

network of large exposures within the global banking system. On the basis of supervisory

reporting of large bilateral exposures we construct the arguably most comprehensive to-

date euro area network of bilateral linkages and combine it with bank balance sheet

information to capture bank-specific characteristics and related (regulatory) solvency and

liquidity constraints.

The CoMap methodology estimates contagion potential due to credit and funding

risks via bilateral linkages. The main objective is to assess the amount of losses and

number of defaults an exogenous shock to a bank (or a group of banks) induces to the

system. In achieving this, the CoMap methodology evaluates first round effects (direct

losses) and subsequent round effects (cascade losses) due to domino defaults and

potential fire sale losses.

We then develop contagion and vulnerability indexes capturing counterparty credit

and funding risks of an exogenous default shock so as to rank banks in terms of

contribution to euro area systemic risk and their degree of fragility, respectively. The

outcome is a practical and quarterly updatable policy tool to map contagion risks

stemming from within and outside the euro area banking system. Overall, the paper

provides unique insights on the interplay of banks’ characteristics and the topology of the

euro area interbank network.

Specifically, the methodology allows for taking a more granular, heterogeneous and

holistic approach to the euro area banking system’s study of contagion risk. Thus, we

model 199 consolidated banking groups (of which 90 from the euro area) in Q3-2017

tracking among them debt, equity, derivative and off-balance sheet exposures larger than

10% of a bank’s eligible capital. We then model banks’ heterogeneity by calibrating the

model’s parameters using exposure-specific information on collateral pledged and

maturity structure as well as bank-specific pool of HQLA and non-HQLA assets and

capital requirements. Overall, the large exposures dataset covers on average 90% of euro

4

area banks’ RWAs vis-à-vis credit institutions for a total amount of EUR 1.4 trillion and

EUR 680 billion, respectively in gross and RWA terms.

We furthermore calculate the contribution of amplification effects (beyond the initial

loss) to the overall losses induced by a bank’s default or distress (amplification ratio), and

we derive a sacrifice ratio indicator assessing the cost-return trade off of a bank-bailout.

Finally, we illustrate how our framework can be used to run counterfactual simulations

showing how contagion risk can be reduced by fine-tuning prudential capital and

liquidity measures.

Key findings highlight that the degree of bank-specific contagion and vulnerability

depends on network specific tipping points affecting directly the magnitude of

amplification effects. It follows that the identification of such tipping points and their

determinants is the essence of an effective micro and macro prudential supervision.

Moreover, we bring evidence that in isolation and with linear variations, bank-specific

characteristics seem to play a less relevant role than the network structure, whereas what

really matters comes from their non-linear interaction, for which both are equally

important. In a variety of tests, heterogeneity in the magnitude of bilateral exposures and

of bank-specific parameters is detected as a key driver of the total number of defaults in

the system. We also show that international spillovers (also coming from non-euro area

banks) are an important channel of contagion for the euro area financial system.

Overall, we think that the CoMap methodology combined with the large exposures

dataset may help enhance our understanding of how contagion within the euro area

interbank network may propagate and be amplified by the actual heterogeneous

characteristics of the agents and the topologic features of the network. It also provides a

practical monitoring toolkit for the regular surveillance and assessment of contagion risk

within the euro area interbank network.

5

“The financial crisis really was a stress test for the men

and the women in the middle of it. We lived by

moments of terror. We endured seemingly endless

stretches when global finance was on the edge of

collapse, when we had to make monumental decisions

in a fog of uncertainty, when our options all looked

dismal but we still had to choose” (Geithner, 2014: 19).

1. Introduction

The collapse of Lehman Brothers has been the defining event of the Great Financial

Crisis of 2007-2008. While the size of its balance sheet alone did not foreshadow the

sequence of events that followed, surely, the uncertainty stemming from its default left

market participants in panic of a wide-spread contagion. Regulators with limited

information about its degree of interconnectedness and bilateral exposures faced the true

dilemma: let it fail or save it. Lehman was allowed to default and the counterfactual

outcome would be debated for years to come. In the last decade, significant progress has

been made in studying the growing interconnectedness of the global financial system and

how shocks are amplified or mitigated depending on the network topology and the

heterogeneity of the agents. However, up to now, uncertainty surrounding the network

due to the lack of available information still represents the major challenge policy makers

face in order to assess the potential cascading effects of such an event. This is the

fundamental question, as highlighted in the incipit of this paper, policy makers and

regulators must answer and be prepared for in case such an adverse event takes place.

Motivated by this question, the systemic risk literature has evolved along two tracks.

The first group of studies try to get around the problem of limited information by relying

on market data such as Acharya et al. (2012, 2017), Billio et al. (2012) and Diebold and

Yilmaz (2014) among others. These market data-based studies allow for capturing

financial institutions’ interconnectedness and to build systemic risk indexes in real time

by exploiting high frequency information on co-movements of stock prices or CDS

spreads. Nevertheless, the interpretation and identification of the underlying mechanism

generating the co-movements may be difficult (Glasserman and Young, 2016). Moreover,

the VAR approaches used to estimate variance decomposition for the forecast errors

suffer from high-dimensionality problems which limits the analysis to small samples of

banks (Alter and Beyer, 2013; Diebold and Yilmaz (2009, 2012). Only recently, Demirer,

6

et al. (2017), Basu et al. (2017), Moratis and Sakellaris (2017) manage to estimate a high-

dimensional network using LASSO methods or Bayesian VARX models. Although

recent innovations in estimation techniques has allowed to increase the sample of banks,

these approaches still cover only a fraction of the banking system, as information on CDS

and stock prices is limited to listed companies. Furthermore, this branch of the literature

does not allow to directly model the interplay of prudential regulations and systemic risk

since the former are only implicitly captured in the degree of co-movements of bank

market prices.

Another stream of the interconnectedness literature hence exploits bilateral

exposures and uses bank balance-sheet based methodologies1. This approach allows for

studying the underlying mechanism of systemic risk formation and contagion stemming

from concrete features of the network, the heterogeneity of the agents, the sources of risk,

and their interplay. In general, balance sheet based studies have tended to focus on a few

specific features so as to better disentangle the path of contagion and amplifications

effects due to e.g. credit risk (Eisenberg and Noe, 2001; Rogers and Veraart, 2013),

funding risks (Gai and Kapadia, 2010; Gai et al. 2011), cross-holdings of assets and fire

sales (Espinoza-Sole 2010; Caballero and Simsek, 2013; Caccioli et al. 2014; Cont and

Schaanning, 2017), as well as from multi-layer networks (Bargigli et al., 2015; Kok and

Montagna, 2016). Overall, these approaches are more theory-based than empirical since

they aim at providing insights on the properties of the network and their implications for

financial stability than actually construct contagion and vulnerability indexes for a

systemic risk assessment as in the market-based approaches.

This is due, among other things, to the lack of availability of a complete set of

bilateral exposures which undermines the accuracy of such systemic risk indicators. In

this respect, most of the empirical literature tends to focus on specific market segment,

overnight or repo markets, or they are country-specific such as studies on the Austrian,

German, Dutch and Italian interbank market (Purh et al. 2012; Craig and von Peter, 2014;

Craig et al. 2014; Veld and Van Lelyveld, 2014; Bargigli et al., 2015). Other studies try

to compensate for the lack of complete network data by imputing missing bilateral

1 See Hüser (2015) for a summary of the literature.

7

linkages based on maximum entropy solution as in Sheldon and Maurer (1998), on

minimum entropy methods (Degryse and Nyguyen, 2007; Elsinger et al. 2006; Upper,

2011), on relative entropy (Van Lelyveld and Liedorp, 2006), or by generating random

networks consistent with partial information (Halaj and Kok, 2013; Anand et al., 2014).

Overall, as emphasized by Glasserman and Young (2016) empirical work in this field

was limited by the confidentiality of interbank transactions and the incomplete set of

information on bilateral exposures. Moreover, these studies focused on rather standard

network measures such as degree centrality, eigenvector centrality and pagerank

algorithms to assess financial system vulnerabilities and systemic importance of banks

Additionally, as access to confidential supervisory data is granted at national level,

most empirical analyses tend to be country-specific. This resulted in the lack of a

comprehensive analysis of cross-border financial exposures, thereby missing bi-

directional linkages with institutions outside a country’s jurisdiction. To a certain extent,

Garratt et al. (2011) and Espinoza-Vega and Sole (2010) overcame this challenge by

using aggregate-level International Consolidated Banking Statistics database from BIS to

assess the cross-border credit and funding risks of a banking system’s default on another

country’s banking system. However, neither study includes bank and exposure level

information thereby ignoring the added value that a specific distribution of exposures and

bank-specific characteristics may bring to the overall stability of the system.

Against this background, this paper aims at overcoming some of the data and

modelling gaps in the interconnectedness literature by studying the degree of contagion

and vulnerability of euro area significant institutions within the global banking system. In

overcoming this challenge, we exploit the actual topology of the euro area interbank

network of large exposures and account for the heterogeneous characteristics of

individual banks via a set of bank and exposure-specific parameters retrieved and

calibrated on ECB supervisory data. This comprehensive data infrastructure allows us to

build a detailed modelling framework capturing the specificities of prudential regulations

such as minimum capital requirements, macroprudential capital buffers, the liquidity

coverage ratio and large exposure limits and their interplay with credit, funding and fire

sales risks.

8

We contribute to the literature in various directions. First, we construct contagion

and vulnerability indexes assessing the systemic footprints of banks. These model-based

estimates allow us to conduct welfare analysis trading off systemic losses due to bank

failures and the cost of policy interventions. Second, we provide a calibration benchmark

for parameters capturing three types of risks: credit, liquidity and fire sales. Third, we

bring evidence that liquidity risk is a major source of default in the interbank network of

large exposures. Fourth, we perform stress test scenarios to assess resilience of the

network structure to wide macro shocks. Fifth, we perform sensitivity analyses to

changes in model parameters so as to assess the non-linear effects derived by the

interplay of network structure and banks’ characteristics. Finally, we provide

counterfactual exercises of prudential measures and their possible usages in reducing the

vulnerability of the network.

We find that tipping points shifting the euro area banking system from a less

vulnerable state to a highly vulnerable state are a non-linear function of the combination

of network structures and bank-specific characteristics. Hence, policies aiming at

reducing systemic risk externalities related to interconnectedness should focus on

increasing the resilience of weak nodes in the system, thereby curbing potential

amplification effects due to cascade defaults, and on reshaping the network structure in

order to set a ceiling to potential losses. Unless systemic risk externalities are internalized

by each bank in the network, bank recapitalizations may be still convenient from a cost-

return trade-off of a global or European central planner. It follows that international

cooperation is essential to limit the ex-ante risk and reduce the ex-post system-wide

losses.

The remainder of the paper is organized as follows. Section 2 presents the data

infrastructure and illustrates the topology of the euro area interbank network of large

exposures. Section 3 details the Contagion Mapping (CoMap) methodology and provides

insights on the calibration of the model parameters. Section 4 discusses the results based

on the contagion and vulnerability indicators and performs sensitivity analysis to assess

the interplay of bank-specific characteristics and the network structure. Section 5 derives

policy implications from a macroprudential supervisor’s perspective via fine-tuning

prudential measures based on counterfactual exercises. The last section concludes.

9

2. Data

The Great Financial Crisis (GFC) of 2008 has led to a rethinking and strengthening of

banking and financial regulation worldwide. The result was the Basel III standards, the

new legal framework aiming at shaping a safer financial system. Macro and micro

prudential regulatory requirements enhancing banks’ capital and liquidity standards as

well as defining leverage and large exposure limits were the actual outcome of this

process.2 In this regard, the large exposures regulation has been developed as a tool for

limiting the maximum loss a bank could incur in the event of a sudden counterparty

failure so as to complement the existing risk-based capital framework (pillar 2) and better

deal with micro-prudential and concentration risks (BIS, 2014).3 Accordingly, banks are

required to report to prudential authorities detailed information about their largest

exposures.

The focus and novelty of this paper is to exploit the microprudential supervisory

framework of large exposures with the aim of constructing the euro area interbank

network to enable us to track the systemic (contagion) risk embedded in the euro area

banking system. The large exposure reporting represents, to our knowledge, the most

comprehensive and up-to-date (on a quarterly basis) dataset capturing granular bank and

exposure level-information of the euro area banking system vis-à-vis entities located

worldwide covering all economic sectors: credit institutions (CIs), financial corporation

(FCs), non-financial corporations (NFCs), general governments (GGs), central banks

(CBs) and households (HHs). In this paper, however, we focus primarily on the

exposures vis-à-vis other credit institutions i.e. the interbank network of large exposures.

Nevertheless, there exist several barriers to utilizing these supervisory data in network

analysis. Because of the confidential nature of this data, the access is generally restricted

to banking supervisors and central banks. However, even for those with access to these

reports, transforming raw data into a suitable format for network analysis is a laborious

task with many challenges. ECB is in a unique position where the supervisory data from

2 This new set of rules was incorporated with the Capital Requirement Regulation (CRR) into EU law,

which from January 2014 applies. This date also coincides with banks’ reporting requirements to National

Central Authorities (NCAs) and to the European Banking Authority (EBA), which ultimately transmits

these large exposures data for monitoring purposes to the Single Supervisory Mechanism (SSM). 3 The large exposure limit is set at 25% of a bank’s eligible capital or 15% for exposures among GSIBs.

10

member states are centrally accessible for monitoring purposes. While this wealth of

information promises high potential, it is a colossal undertaking to reconcile this data

across many jurisdictions and set-up a euro area banking network of large exposures for a

comprehensive systemic risk assessment.

2.1 Large Exposures

An exposure is considered a “large exposure” when, before applying credit risk

mitigations and exemptions, it is equal or higher than 10% of an institution’s eligible

capital vis-à-vis a single client or a group of connected clients (CRR, art. 392).4

Moreover, institutions that report FINREP supervisory data are also requested to report

large exposures information with a value above or equal to EUR 300 million. Therefore

the data sample coverage is very comprehensive and captures almost EUR 13.5 trillion of

gross exposures in Q3 2017 (our reference date), more than 50% of euro area credit

institutions’ total assets. In risk-weighted terms the coverage is smaller but still

comprehensive, capturing almost 40% of the total RWAs of euro area banks. However, in

terms of studying the euro area interbank network, which is the subject of our paper, the

large exposures sample captures 90% of euro area banks’ RWAs vis-à-vis credit

institutions. This extensive coverage provides us with confidence that we can reliably

model euro area banks’ degree of interconnectedness and their contribution to cross-

sectional systemic risk.

It is also notable that the large exposure data go well beyond the standard unsecured

interbank transactions typically covered in many interbank network studies. In fact, in the

supervisory reporting a large exposure is defined as any direct and indirect debt,

derivative, equity, and off-balance sheet exposure that complies with the reporting

threshold.5 In this regard, a key feature of the regulation is that the counterparty may be

identified not only as an individual client, but also as a group of connected clients (CRR,

4 Eligible capital is defined as the sum of tier 1 capital plus one-third or less of tier 2 capital (CRR, art. 4:

71). 5 Direct exposures refer to exposures on “immediate borrower” basis, while an indirect exposure, according

to article 403 of CRR, is an exposure to a client guaranteed by a third party, or secured by collateral issued

by a third party. Moreover, according to article 399 of CRR exposures arising from credit-linked notes shall

also be reported as indirect exposures.

11

art. 4:1:39).6 The latter refers to the fact that the reporting institution needs to assess and

take into account - not only direct and indirect risks - but also possible domino effects

and negative externalities from funding shortfalls due to control relationships and

economic interdependencies (EBA/GL/2017/15). This is a highly relevant feature and an

added value of the data because large exposures allow us to capture the full risk spectrum

and not only the actual amount of exposures at risk. Moreover, in achieving this, a

standardized evaluation method is applied so that the totaling final exposure amount is

reliably comparable across countries and reporting institutions7.

2.2 Dataset

This subset of large exposures data captures almost completely credit and funding risks

of euro area SIs among themselves, and credit risks of EA SIs vis-à-vis non-euro area

banks. However, the large exposures dataset does not capture euro area SIs’ funding risks

from non-euro area banks. In order to tackle this information gap, we retrieve data on the

10 largest funding sources of euro area SIs by using another COREP supervisory

template defined as concentration of funding by counterparty or C.67.8 These 10 largest

funding sources may come from central banks, governments, credit institutions, and

corporates. Regarding, our sample of counterparties we find 41 funding exposures from

non-euro area banks towards euro area SIs. We incorporate them into the large exposures

dataset matching them with the gross exposures before exemptions and credit risk

mitigations. Next, we consolidate the large exposures data from euro area SIs that are

euro area-based subsidiaries of non-euro area banks with these 41 funding sources from

non-euro area banks. We then drop intra-group exposures and set a 50 million threshold

for exposure before credit risk mitigations (but after exemptions) to clean-up the network

from negligible edges.

6 EBA’s guidelines on connected clients - final report - further detail Art. 4(1)(39) emphasizing that

whether financial difficulties or a failure of a client would not lead to funding or repayment difficulties for

another client, these clients do not need to be considered a single risk (e.g. where the client can easily find a

replacement for the other client). Moreover, these guidelines point out that a reporting institution should

investigate all economic dependencies for which the sum of all exposures to one individual client exceeds

5% of Tier 1 capital. 7 Details on the construction of the dataset are provided in the appendix. 8 A minimum threshold of 1 percent of total liabilities applies either as a single creditor or a group of

connected clients.

12

On top of this, for modelling purposes, we retrieve from other COREP supervisory

templates euro area SI’s RWAs (C.02.10), the total capital base (C.01.10), Tier 1

(C.01.15), CET1 (C.01.20), pillar 2 capital requirements (C.03.00), and capital buffers

(C.06.01, and ECB website).9 In this respect, bank-specific combined capital buffers

were constructed following regulatory capital requirements: banks’ minimum capital

requirement (MC), pillar 2 requirements (P2R), the capital conservation buffer (CCoB),

the OSII, GSII and the systemic risk buffers (SRB), as well as the prevailing

countercyclical capital buffer requirements (CCyB).10

Regarding the international

Globally Systemic Important Institutions, the GSII buffers were retrieved from the

Financial Stability Board 2017 list.11

Moreover, in order to calibrate the model we also

retrieve from COREP templates the maturity buckets of large exposures (template C.30),

banks’ HQLAs (C.72.00.a.10), the liquidity buffer and net liquidity outflow, respectively

numerator and denominator of the liquidity coverage ratio (template C.76.00.a.10/20) as

well as from FINREP templates banks’ total assets and their components (F.1.1.380).12

Whereas, for the international banks we match our data set to Bankscope for bank

characteristics, and when missing, we use the most recent annual consolidated financial

report.13

In the end, we limit our data set to banks that has a complete set of information

on the above described metrics. This yields our main dataset of 199 consolidated banking

groups or nodes, whose total assets amount up to EUR 74 trillion, approximately 6.6

times the GDP of the euro area.

Overall, this brings us to a total number of large exposures equal to 1.734, and a total

gross amount of EUR 1.38 trillion, or EUR 675 billion of risk-weighted assets. This sub-

set of SIs’ large exposures to credit institutions cover almost 80% of the total gross

9 ECB Website: http://www.ecb.europa.eu/pub/fsr/html/measures.en.html 10 Minimum capital requirements are defined as follows: 4.5% of CET1, 6% of TIER1, and 8.0% of own

funds that is the sum of 4.5% CET1 + 1.5% AT1 + 2% T2. Pillar 2 measures are included on top of

minimum capital requirements and are mandatory only within the EU. CCoB may take value of 1.875%

CET1 if the country is still in a transitional period, or 2.5% CET1 if fully-loaded. 11 Financial Stability Board’s 2017 list of global systemically important banks: http://www.fsb.org/wp-

content/uploads/P211117-1.pdf 12 Tangible assets are calculated as total assets minus intangible assets (300), tax assets (330), other assets

(360), and non-current assets and disposal groups classified as held for sale (370). 13 In few cases for international banks, variables were approximated using a balance sheet-based

methodology (discussed in the calibration section) using as reference value the average of the sample.

13

amount. Furthermore, the selected sample of counterparties guarantees a complete

coverage of LEI codes, country of domicile and sector of belonging.

Table 1 presents the summary statistics of the interbank network of large exposures in

Q3 2017. It consists of 179 counterparties and 101 reporting institutions, for a total of

1264 exposures (or edges), of which, almost 90% (1185) is reported by euro area-based

banking groups, and the remaining 10% (79) from international banks. The latter

provides a partial picture of the euro area banks’ funding risk related to non-euro area

creditors. On the contrary, euro area banks’ credit risk is captured in its entirety, and it is

distributed almost equally between euro area and extra-euro area counterparties,

amounting to respectively 613 and 651 in terms of number of exposures (edges) or EUR

431 billion and EUR 432 billion (gross amount minus exemptions). In comparison, gross

exposures before exemptions and CRM (gross amount) amount to EUR 1.13 trillion,

while after exemptions and CRM, exposures amount up to EUR 623 billion (net amount).

No bilateral-linkages among extra-euro area banks are captured in this study. Therefore,

to the best of our knowledge, in terms of coverage this dataset represents one of the most

comprehensive attempts to study euro area systemic risks by means of granular bank and

exposure level-information.

Table 1: Interbank Network of Large Exposures

Note: Amounts are expressed in billions of euros. Outstanding amounts as of Q3 2017. Gross amount

minus exemptions is the reference metrics of this study. A 50 million threshold to exposures before credit

risk mitigation was applied. Exemptions are those amounts which are exempted from the large exposure

calculation, whereas credit risk mitigations refer to the amounts adjusted for risk weights.

Data Sample Total Euro Area non-Euro Area

Entities

Consolidated Banking Groups 199 90 109

Counterparties 179 69 110

Reporting 101 84 17

Number of Exposures

From 1264 1185 79

To 1264 613 651

Total Exposures Amount

Gross Amount 1126 639 487

- Exemptions 263 208 55

Gross Amount minus Exemptions 863 431 432

- Credit Risk Mitigations 240 165 75

Net Amount 623 266 357

14

2.3 Network Topology

We are now in the position to plot the euro area interbank network of large exposures.

For a graphical interpretation of the edges’ directionality, we drop from the interbank

network the 79 funding linkages from international banks. This allows us to split each

chart into two concentric circles. An inner-circle composed only of euro area banks’

credit and funding exposures among themselves (EA), and an outer-circle reflecting the

international dimension of euro area banks’ credit exposures towards non-euro area banks

(INT). Therefore all edges linking the inner circle to the outer circle are outward oriented.

Figure 1 displays the network according to two mirror images, capturing respectively

the size of exposures in Euro Billions - panel (a) - and in % of lender’s capital - panel (b).

In both panels the edges take the color from the node borrowing the fund, i.e. the amount

international banks are borrowing from euro area banks. Given the lack of credit

exposures from international banks to euro area banks, for comparative purposes, we

assign as a node’s size the weighted in-degree, that is, the sum of incoming exposures.

Therefore, the node’s size captures the relative size of each bank’s funding exposure.

On the one hand, panel (a) identifies which international banking system is the most

interconnected with the euro area banking system. For instance, euro area banks appear to

have few exposures but sizeable (thick) towards Chinese banks, many but relatively small

exposures to Swiss banks, and many and sizeable exposures to US and UK banks.

Overall, international spillovers seem to be an important channel of contagion to the euro

area banking system. On the other hand, panel (b) of figure 1, which presents the

interlinkages in percentage of the lenders’ capital shows that the node size of

international banks becomes slightly smaller due to the fact that euro area medium and

large-sized banks tend to lend more to international banks than small domestic banks,

which by comparison tend to lend more to banks within the euro area. In this respect,

even if not clearly visible, small-medium banks tend to have relatively fewer cross-border

large exposures both within the inner circle and with the outer circle, implying that the

potential for cross-country spillovers is likely to mostly pass-through the major country

hubs. Overall, the euro area interbank network of large exposures can be characterized by

a core-periphery network structure. This feature also results in a relatively sparse

network. In fact, only 6.3% of all possible links are present.

15

Figure 1. Euro Area Interbank Network of Large Exposures - Borrower Perspective

Euro Billions - Panel (a) % of Capital - Panel (b)

Source: COREP C.27-C.28.

Note: The size of the nodes captures the weighted in-degree of interconnectedness. The colours of nodes

are clustered by country of origin, the thickness of the flows summarizes the value of the exposures in EUR

billions and percentage of eligible capital, respectively. The colour of the flows refers to the target of the

node’s colour capturing respectively the borrower perspective.

3. Contagion Mapping (CoMap) Methodology

This paper relies primarily on a balance sheet simulation approach to map contagion. In

addition to demonstrating the architecture of banking networks through bilateral linkages,

such an approach also allows us to quantify systemic losses and determine channels of

contagion by assuming hypothetical failures in the network. The emphasis on granularity

in establishing bilateral connections applies equally in modeling contagion. By

incorporating model parameters which are calibrated based on bank-specific - and to the

extent possible exposure-specific - data allows us to simulate a contagion model that

provides a fairly accurate picture of the reality.

3.1 Modelling Framework

Our Contagion Mapping model (CoMap) is essentially a variant of the Eisenberg and

Noe (2001) framework. This framework has been at the center of many studies in the

financial networks literature. Our starting point is a simple interbank exposure model

EA US UK CN CH SE-NO-DK BR-IN-RU JP TR ROW

EA

INT

16

with both credit and funding shocks.14

Credit shocks capture the impact of a bank

defaulting on its liabilities to other banks. Funding shocks, on the other hand, represent

how a bank’s withdrawal of funding from other banks forces them to deleverage by

selling assets at a discount (fire sale). Triggering a distress event (single or multiple bank

failures) reveals the cascade effects and propagation channels transmitted through these

solvency and liquidity channels. In order to achieve a more realistic setting, we enrich

this simple framework with a series of new features that reflect heterogeneity across

banks, one of the novelties of this paper. Specifically, we model the effects of: (i) bank-

specific default thresholds, such as minimum capital requirements and capital buffers; (ii)

changes to the network structure via large exposure limits; (iii) variations in exposures at

risk (loss-given-default); (iv) maturity structure of bank funding; (v) market risk linked to

a bank’s business model captured by the amount of financial and HQLA assets on a

bank’s balance sheet; (vi) changes in bank-specific LCR ratio due to adjustments in the

liquidity buffer and/or the net liquidity outflows. As a result, this comprehensive

modelling framework is able to capture the risk-return trade-off a bank faces between

holding HQLA and non-HQLA financial assets and allows for assessing both solvency

and liquidity risk while accounting for bank-specific parameters. Hence, it incorporates

(vii) liquidity constraint on the amount of assets available for sale allowing a bank to

default because of being illiquid. These seven distinctive features are jointly modelled in

our framework.

The initial set-up of our model, while closely following Espinosa-Vega and Sole

(2010), expands the scope beyond interbank loans to capture all interbank claims. This is

reflected in the stylized balance sheet identity of bank i as follows:

∑ ∑ 𝑥𝑖𝑗𝑘

𝑘𝑗+ 𝑎𝑖 = 𝑐𝑖 + ⅆ𝑖 + 𝑏𝑖 +∑ 𝑥𝑗𝑖

𝑗 (1)

where 𝑥𝑖𝑗𝑘 stands for bank i's claims of type k on bank j, 𝑎𝑖 stands for other assets, 𝑐𝑖

stands for capital, 𝑏𝑖 are wholesale funding (excluding interbank transactions), ⅆ𝑖 stands

for deposits, and 𝑥𝑗𝑖𝑘 stands for bank i’s total obligations vis-à-vis bank j, or conversely,

14 Espinosa-Vega and Sole (2010) illustrate the workings of such a model with the use of aggregated data.

17

bank j’s claims on bank i. 𝒵 is the complete set of all banks in the network with a total of

N number of banks.

Next, we introduce the key elements of our baseline model that will be used as a

reference framework in the remainder of this paper.

3.1.1 Credit Shock

Credit shock captures the impact of a bank or a group of banks defaulting on their

obligations to other banks. As a result, a bank incurs losses on a share of its claims

depending on the nature and counterparty of its exposures. Other studies have assumed

uniform loss-given default rates, be it at entity level or for the entire network.15

In

practice, different claims may have different recovery rates. For example, the recovery

rates from equity stakes and debt claims can vary. We introduce exposure-specific loss-

given default rates to reflect the precise risk mitigation and collateralization a bank has

accounted for its claims vis-à-vis each counterparty. In response to a subset of banks,

𝒴 ⊂ 𝒵, defaulting on their obligations, bank i’s losses are summed across all banks 𝑗 ∈ 𝒴

and claim types 𝑘 using exposure-specific loss-given default rates, 𝜆𝑖𝑗𝑘 , corresponding to

its claim of type k on bank j, 𝑥𝑖𝑗𝑘 :

∑ ∑ 𝜆𝑖𝑗𝑘 𝑥𝑖𝑗

𝑘

𝑘, 𝑤ℎ𝑒𝑟𝑒 𝜆𝑖𝑗

𝑘

𝑗∈𝒴∈ [0,1] 𝑎𝑛ⅆ 𝑖 ∉ 𝒴 (2)

The total losses are absorbed by bank i's capital while the size of its assets is reduced

by the same amount.

∑ ∑ 𝑥𝑖𝑗𝑘

𝑘𝑗∈𝒵\𝒴+ [𝑎𝑖 +∑ ∑ (1 − 𝜆𝑖𝑗

𝑘 )𝑥𝑖𝑗𝑘

𝑘𝑗∈𝒴]

= [𝑐𝑖 − ∑ ∑ 𝜆𝑖𝑗𝑘 𝑥𝑖𝑗

𝑘

𝑘𝑗∈𝒴] + ⅆ𝑖 + 𝑏𝑖 +∑ 𝑥𝑗𝑖

𝑗 (3)

As a result, bank i's balance sheet shrinks, with lower capital, 𝑐𝑖′, reflecting the losses.

The recouped portion of its claims are commingled with other assets, 𝑎𝑖′.

15 See for instance: Battiston et al. (2012), Cifuentes et al. (2005), Cont et al. (2010), Espinosa-Vega and

Sole (2010) and Rogers and Veraart (2013).

18

∑ ∑ 𝑥𝑖𝑗𝑘

𝑘𝑗∈𝒵\𝒴+ 𝑎𝑖

′ = 𝑐𝑖′ + ⅆ𝑖 + 𝑏𝑖 +∑ 𝑥𝑗𝑖

𝑗 (4)

Figure 2 illustrates the transmission of credit shock via bilateral linkages on bank i's

balance sheet.

Figure 2. Impact of Credit Shock on Bank i’s Balance Sheet

before credit shock credit shock after credit shock

3.1.2 Funding Shock

Funding shock represents how a bank’s withdrawal of funding from other banks forces

them to deleverage by selling assets at a discount (fire sale). Typically, an assumption is

made about the share of short-term funding that cannot be rolled over and the haircut rate

that must be applied to fire sale of assets to meet the immediate liquidity needs. This

would result in losses on the trading book, which would then be absorbed by the capital

base. We introduce bank-specific funding shortfall rate, 𝜌𝑖, reflecting precisely the

maturity structure of bank i's wholesale funding. In response to a subset of banks

defaulting (getting into distress), 𝒴 ⊂ 𝒵, and thereby withdrawing funding from other

counterparties, bank i faces funding shortfall summed across all banks 𝑗 ∈ 𝒴 using its

specific funding shortfall rate, 𝜌𝑖:

∑ 𝜌𝑖𝑥𝑗𝑖 , 𝑤ℎ𝑒𝑟𝑒 𝜌𝑖𝑗∈𝒴

∈ [0,1] (5)

19

We introduce to the model banks’ ability to hold liquidity surplus, which can be used

to absorb these shortfalls, at least partially. In order to mitigate banks’ short-term funding

risk, regulators have imposed liquidity coverage ratios (LCR) to ensure that banks have

sufficient high-quality liquid assets (HQLA) to cover liquidity shortages. In practice, for

immediate liquidity needs, banks can pledge HQLA as collateral to the central bank for

overnight borrowing. From a modeling perspective, this implies that bank i can offset

funding shortfall with the new credit line up to its liquidity surplus, 𝛾𝑖:

𝑚𝑖𝑛 {𝛾𝑖 ,∑ 𝜌𝑖𝑥𝑗𝑖𝑗∈𝒴

} (6)

with the remaining liquidity shortage computed as:

𝑚𝑎𝑥 {0,∑ 𝜌𝑖𝑥𝑗𝑖𝑗∈𝒴

− 𝛾𝑖} (7)

In our model, a bank is pushed toward a fire sale when it has exhausted emergency

credit lines from the central bank, that is, if the remaining liquidity shortage (7)

emanating from the funding shock is strictly positive. At this point, we introduce a

constraint, 𝜃𝑖, on the amount of remaining assets available to the bank to sell. This

constraint sets an upper threshold to how much of the remaining liquidity shortage can be

sustained with the fire sale proceeds after accounting for haircuts proportional to a

discount rate, 𝛿𝑖. As a result, the deleveraging amounts to the sale of assets is equivalent

to:

𝑚𝑖𝑛 {1

1 − 𝛿𝑖𝑚𝑎𝑥 {0,∑ 𝜌𝑖𝑥𝑗𝑖

𝑗∈𝒴− 𝛾𝑖} , 𝜃𝑖} , 𝑤ℎ𝑒𝑟𝑒 𝛿𝑖 ∈ [0,1] (8)

As in credit shock, the losses due to the fire sale are absorbed fully by bank i's

capital. The other liabilities of the bank decline by the amount of funding shortfall that

couldn’t be replenished by central bank loans. The sum of the two declines are matched

by the contraction on bank’s assets due to fire sales.

20

∑ ∑ 𝑥𝑖𝑗𝑘

𝑘𝑗+ [𝑎𝑖 −𝑚𝑖𝑛 {

1

1 − 𝛿𝑖𝑚𝑎𝑥 {0,∑ 𝜌𝑖𝑥𝑗𝑖

𝑗∈𝒴− 𝛾𝑖} , 𝜃𝑖}]

= [𝑐𝑖 − 𝛿𝑖𝑚𝑖𝑛 {1

1 − 𝛿𝑖𝑚𝑎𝑥 {0,∑ 𝜌𝑖𝑥𝑗𝑖

𝑗∈𝒴− 𝛾𝑖} , 𝜃𝑖}] + ⅆ𝑖

+ [𝑏𝑖 +𝑚𝑖𝑛 {𝛾𝑖 ,∑ 𝜌𝑖𝑥𝑗𝑖𝑗∈𝒴

}] + [∑ 𝑥𝑗𝑖𝑗∈𝒵\𝒴

+∑ (1 − 𝜌𝑖)𝑥𝑗𝑖𝑗∈𝒴

] (9)

Overall, the balance sheet of the bank can potentially shrink by a larger factor than

the associated capital losses in contrast with the credit shock. On the liabilities side, there

is a shift in wholesale funding from other banks to the central bank, as well as an overall

decline.

∑ ∑ 𝑥𝑖𝑗𝑘

𝑘𝑗+ 𝑎𝑖

′′ = 𝑐𝑖′′ + ⅆ𝑖 + 𝑏𝑖

′′ + [∑ 𝑥𝑗𝑖𝑗∈𝒵\𝒴

+∑ (1 − 𝜌𝑖)𝑥𝑗𝑖𝑗∈𝒴

] (10)

Figures 3 and 4 illustrate the transmission of funding shock via bilateral linkages on

bank i's balance sheet, when the liquidity surplus is sufficient to meet funding shortfall

and when it is insufficient, respectively.

Figure 3. Impact of funding shock on bank i’s balance sheet with sufficient buffer

before funding shock funding shock after funding shock

21

Figure 4. Impact of funding shock on bank i’s balance sheet with insufficient buffer

before funding shock funding shock after funding shock

3.1.3 Simultaneous Credit and Funding Shocks

While it is helpful to consider credit and funding shocks in isolation, when a bank or a

group of banks are in distress, they are likely to default on their obligations and shore up

liquidity by withdrawing funding simultaneously. Therefore, we combine the impact of

both shocks on bank i's balance sheet to capture the full impact of a distress event.

[𝑎𝑖 +∑ ∑ (1 − 𝜆𝑖𝑗𝑘 )𝑥𝑖𝑗

𝑘

𝑘𝑗∈𝒴−𝑚𝑖𝑛 {

1

1 − 𝛿𝑖𝑚𝑎𝑥 {0,∑ 𝜌𝑖𝑥𝑗𝑖

𝑗∈𝒴− 𝛾𝑖} , 𝜃𝑖}]

+∑ ∑ 𝑥𝑖𝑗𝑘

𝑘𝑗∈𝒵\𝒴

= [𝑐𝑖 − ∑ ∑ 𝜆𝑖𝑗𝑘 𝑥𝑖𝑗

𝑘

𝑘𝑗∈𝒴− 𝛿𝑖𝑚𝑖𝑛 {

1

1 − 𝛿𝑖𝑚𝑎𝑥 {0,∑ 𝜌𝑖𝑥𝑗𝑖

𝑗∈𝒴− 𝛾𝑖} , 𝜃𝑖}] + ⅆ𝑖

+ [𝑏𝑖 +𝑚𝑖𝑛 {𝛾𝑖 ,∑ 𝜌𝑖𝑥𝑗𝑖𝑗∈𝒴

}]

+ [∑ 𝑥𝑗𝑖𝑗∈𝒵\𝒴

+∑ (1 − 𝜌𝑖)𝑥𝑗𝑖𝑗∈𝒴

] (11)

22

3.1.4 Default Mechanisms

Up to this point, we focused on how credit and funding shocks are transmitted to a

bank’s balance sheet. While credit shocks translate directly to weakening of a bank’s

capital, funding shocks lead to depletion of its liquidity and via fire sales to capital losses.

Now, we define at what level these losses result in a severe distress for a bank triggering

its default.

In a distress event, the capital of exposed counterparties, such as bank i, must absorb

the losses on impact. Then, bank i becomes insolvent if its capital falls below a certain

threshold 𝑐𝑖𝑑, which may be defined as the bank’s minimum capital requirements with or

without capital buffers. In other words, bank i is said to fail if its capital surplus (𝑐𝑖 − 𝑐𝑖𝑑)

is insufficient to fully cover the losses:

𝑐𝑖 − 𝑐𝑖𝑑 < ∑ ∑ 𝜆𝑖𝑗

𝑘 𝑥𝑖𝑗𝑘

𝑘𝑗∈𝒴+ 𝛿𝑖𝑚𝑖𝑛 {

1

1 − 𝛿𝑖𝑚𝑎𝑥 {0,∑ 𝜌𝑖𝑥𝑗𝑖

𝑗∈𝒴− 𝛾𝑖} , 𝜃𝑖} (12)

In terms of the impact through the liquidity channel, bank i’s liquidity surplus serves

as the first line of defense. However, the remaining liquidity shortages might require a

large-scale fire sale operation relative to its financial assets. Having already exhausted its

liquidity surplus, bank i becomes illiquid if its remaining assets are insufficient to match

the liquidity shortage:

𝜃𝑖 < 1

1 − 𝛿𝑖𝑚𝑎𝑥 {0,∑ 𝜌𝑖𝑥𝑗𝑖

𝑗∈𝒴− 𝛾𝑖} (13)

Notably, in our framework, a bank may default contemporaneously via solvency and

liquidity when inequalities (12) and (13) are jointly satisfied. This implies that the

funding shortfall is larger than the funds retrieved from the liquidity surplus and the fire

sale operations, and, at the same time, the cumulated losses incurred via credit losses and

fire sales are larger than the capital surplus.

Bringing the full network of banks into picture, in each simulation the exercise tests

the system for a given bank’s default as depicted in Figure 5. The initial default of bank 1

is triggered by design in order to study the cascade effects and contagion path it causes

through the interbank network. According to this example, the trigger bank is linked to

23

bank 2 and bank 4 via large exposures, respectively 𝑥12 and 𝑥14. The initial shock

determines the subsequent bank 2’s default and losses to bank 4 via credit and funding

risks. Hence, the exercise continues to the second round since there is at least one

additional failure in response to the initial exogenous shock. In this round, banks’ losses

are cumulated in calculation of their distress conditions. Therefore, bank 4’s losses

experienced by bank 2’s default (𝑥24.) in round 2 are summed up with bank 1’s induced

losses in round 1 (𝑥14.). Although, the initial default of bank 1 does not directly induce

bank 4’s default, due to contagion and amplification effects, in round 2 bank 4‘s default

realizes. In turn, bank 4 triggers the default of bank 3 (𝑥43) and produces additional

losses to bank 2 (𝑥42) which already defaulted in round 2. In this respect, the losses

experienced by bank 2 via the large exposure (𝑥42) will not further affect bank 2 since its

surplus of capital above the minimum has been already depleted given bank 1 shock.16

The exercise moves to subsequent rounds if there are additional failures in the system and

stops when there are no other failures.

Figure 5: Contagion Path and Rounds to Defaults

Note: The trigger bank initializes the algorithm, rounds track the path of contagion via internal loops, while

final failures define the convergence of the algorithm.

Source: Inspired by Espinoza-Sole (2010).

16 This is an assumption of the model that may be relaxed depending on whether we want to model the

entire distress induced to the capital base or simply to the capital base above the minimum capital

requirements.

24

3.2 Calibration

The typical approach in the application of balance sheet simulation exercises has been to

use benchmark parameters based on cross-country studies or sectoral averages. Few

studies introduced some improvement by using random drawings from a distribution of

observed values. One of the main contributions of this paper is to model bank-level

heterogeneity with granular exposure and other balance sheet information. In the

following, we describe in detail how we calibrate the bank-specific parameters for the

set-up of our benchmark model.

3.2.1 Loss Given Default

The loss given default (LGD) parameter is calibrated for each bank at exposure level by

calculating the ratio of net exposures to gross exposures. Gross exposures (GE) are

defined as those after deducting defaulted amounts and exemptions from original gross

exposures. Net exposures (NE) refer to the remaining exposures after adjusting gross

exposures for credit risk mitigation measures. In other words, if bank i is lending to

counterparty j, the exposure-specific LGD is defined as in equation (14). Non-reporting

banks in the sample are assumed to have a uniform LGD equal to the average �̅� across all

reporting banks.

𝜆𝑖,𝑗 =𝑁𝐸𝑖,𝑗𝐺𝐸𝑖,𝑗

= 𝐿𝐺𝐷𝑖,𝑗 (14)

On the one hand, panel (a) of Figure 6 presents the distribution of the exposure-

specific loss given default parameters (𝜆𝑖,𝑗). The red line shows the average of the

sample (�̅�) upon which is based the calibration for the non-reporting banks. The average

net exposure amount is 80% of the gross amount after deducting exemptions. Panel (b)

reports the distribution of exemptions across exposures. Both samples are concentrated

respectively on the right and left side of the distribution, though cross-exposure

heterogeneity is visible.

25

Figure 6: Exposure-Specific Loss Given Default Parameter

Panel (a) Panel (b)

Source: COREP Supervisory Data, Template C.28.00.

Note: The LGD parameter is calculated on an exposure basis as the share of the net exposure (after CRM

and exemptions) over the gross exposure amount before taking into account CRM, but after exemptions.

The exemption rate shows the share of exempted amount over the original gross amount before deducting

exemptions and CRM.

3.2.2 Funding Shortfall

Funding shortfall refers to the portion of withdrawn funding that is assumed not to be

rolled over when the bank providing the funding defaults (or gets into distress). It is

calibrated at bank-level using the share of short-term liabilities shorter than 30 days (1

month maturity). The choice of this maturity threshold as baseline calculation is to allow

the funding shortfall to be consistent with the Liquidity Coverage Ratio (LCR) which

assumes a 30-day liquidity distress scenario. However, this assumption may be relaxed

and 𝜌𝑖 can be calibrated on a shorter or longer period depending on the scenario we want

to test.

For each bank, we use exposure level information retrieved from the concentration

of funding template (C.67.00.a) and the large exposure maturity breakdown template

(C.30). The former template allows us to retrieve information on the exposures’ amount

and maturity breakdown on international banks lending to euro area banks. Therefore, as

reported in equation (15), the funding shortfall is calibrated based on the share of

exposures in buckets with maturities of less than 30 days over the total amount of

26

funding, aggregated across all reporting banks for whom bank i is a large exposure

counterpart (Fi). 17

𝜌𝑖 =∑ 𝑥𝑖,𝑗<30𝑑𝑎𝑦𝑠𝑗𝜖𝐹𝑖

∑ 𝑥𝑖,𝑗𝑗𝜖𝐹𝑖

=𝑆ℎ𝑜𝑟𝑡 𝑇𝑒𝑟𝑚 𝐹𝑢𝑛ⅆ𝑖𝑛𝑔

𝑇𝑜𝑡𝑎𝑙 𝐹𝑢𝑛ⅆ𝑖𝑛𝑔 (15)

When no maturity information is available, we use the average maturity to which the

reporting banks having an exposure to bank i are lending at to other banks. Therefore, we

assume that the maturity information of the reporting bank is more accurate than setting

𝜌𝑖 equal to the average of the sample. This approach allows us to increase heterogeneity

in the distribution of the funding shortfall parameter.

Figure 7: Bank-Specific Funding Shortfall Parameter

Source: COREP Supervisory Data, Template C.30 and Template C.67.00.a

Note: Funding shortfall is constructed as short-term funding divided by total funding.

As we see in figure 7, banks’ short term funding as share of total funding is

distributed on the whole range of the maturity breakdown, with banks experiencing an

average of 35% of short term funding over total funding.

3.2.3 Liquidity Surplus

The liquidity surplus is directly derived from the liquidity coverage ratio template

C.72.00a. It consists of the difference between the LCR’s numerator and denominator

since the former, as of 2018, needs to be larger than 100% of the latter (equation 16).

Hence, the liquidity surplus (𝛾𝑖) refers to the stock of HQLAs (𝐿𝐵𝑖) above the net

17 Bank i’s large exposure vis-à-vis bank j can be equally thought of as the amount of funding provided by

bank i to bank j.

27

funding outflows (𝑁𝐿𝑂𝑖) over a 30-day liquidity distress scenario. Figure 8 reports the

surplus as share of banks’ total assets. The average of the sample is close to 5.8% which

is used for approximating the missing (𝛾𝑖) for some international banks. Furthermore, if a

bank is currently facing a transition period to achieve the 100% LCR ratio, whenever

𝑁𝐿𝑂𝑖 > 𝐿𝐵𝑖, to be conservative, we set 𝛾𝑖 = 0.

𝐿𝐶𝑅: 𝐿𝐵𝑖𝑁𝐿𝑂𝑖

> 1 𝑦𝑖𝑒𝑙𝑑𝑠→ 𝐿𝐵𝑖 > 𝑁𝐿𝑂𝑖

𝑦𝑖𝑒𝑙𝑑𝑠→ 𝛾𝑖 ≡ 𝐿𝐵𝑖 − 𝑁𝐿𝑂𝑖 > 0 (16)

Figure 8: Bank-Specific Liquidity Surplus

Source: COREP Supervisory Data, Template C.72.00.a and Bankscope.

Note: Liquidity Surplus (𝜸) is constructed as the difference between the numerator and the denominator of

the liquidity coverage ratio (LCR), i.e. the difference between the stock of HQLAs (LB) and the net

funding Outflows (NFO).

3.2.4 Fire Sale Discount Rate and Pool of Assets

The additional parameters required to simulate the contagion impact of a funding shock is

the rate at which banks are forced to discount their assets as they react to a funding

shortfall by deleveraging. Since, as the described in the previous section, we assume that

the set of HQLA assets is used to cover the liquidity shortfall, and the fire sale stage is

triggered only when it is exhausted, the set of assets available for sale is defined as the

amount of unencumbered non-HQLA assets. This category of assets is retrieved from the

asset encumbrance template F.32.01 which is further broken-down into different asset

classes. In this respect, Equation (17) approximates the discount rate (𝛿𝑖) as the ratio

between the discounted amount of unencumbered non-central bank eligible assets

(𝐷_𝑈𝑁𝐶𝐵𝐸𝐴𝑖) over the total amount of unencumbered non-central bank eligible assets

28

(𝑈𝑁𝐶𝐵𝐸𝐴𝑖), which captures the pool of assets available for sale (𝜃𝑖). Therefore the

𝛿𝑖 coefficient for euro area banks is derived as the weighted average haircut (𝛿�̅�) of each

asset classes 𝐴𝑗: respectively covered bonds (𝛿�̅�𝐵), asset backed securities (𝛿�̅�𝐵𝑆), debt

securities issued by general governments (𝛿�̅�𝐺), debt securities issued by financial

corporations (𝛿�̅�𝐶), debt securities issued by non-financial corporations (𝛿�̅�𝐹𝐶), and

equity instruments (𝛿̅̅̅𝐸). The average haircut (𝛿�̅�) for each asset class is based on the

latest ECB’s guidelines on haircuts.18 Moreover, in order to take into account that the

instruments we are dealing with are non-central bank eligible, we assume that the bottom

threshold for haircuts is the highest haircut for central bank eligible instrument, i.e. 38%.

𝛿𝑖 =∑𝛿�̅�𝐴𝑗𝐴𝑗

=

𝑁

𝑗

𝛿�̅�𝐵𝐶𝐵𝑖 + 𝛿�̅�𝐵𝑆𝐴𝐵𝑆𝑖 + 𝛿�̅�𝐺𝐺𝐺𝑖 + 𝛿�̅�𝐶𝐹𝐶𝑖 + 𝛿�̅�𝐹𝐶𝑁𝐹𝐶𝑖 + 𝛿�̅�𝐸𝑖𝑈𝑁𝐶𝐵𝐸𝐴𝑖

(17)

For international banks for which we lack FINREP template F.32.01, we derive the

discount rate 𝛿𝑖 and the pool of assets available for sale (𝜃𝑖) with a two-step procedure.

First, we regress the balance sheet categories i) assets available for sales, ii) assets held

for trading and iii) HQLA assets for the euro area banks sample as reported in equation

(18) on the numerator and denominator of equation (17).

∑𝛿𝑖,𝑗̅̅ ̅̅ 𝐴𝑖,𝑗

𝑁

𝑗

= 𝑎1𝐹𝐴𝐴𝑆𝑖 + 𝑎2𝐹𝐴𝐻𝑇𝑖 + 𝑎3𝐻𝑄𝐿𝐴𝑖 + 𝑒𝑖 (18)

In this way we obtain three coefficients 𝑎1, 𝑎2, 𝑎3 explaining the contribution of each

asset class for both dependent variables. As we can see from Table 2, the first two

coefficients are statistically significant at 1% and the model shows a reliable goodness of

fit, respectively 89% and 86% for the numerator and denominator of equation (17). Next,

we retrieve from Bankscope the very same balance sheet categories for which we have a

statistically significant coefficient, i.e., financial assets available for sale and financial

assets held for trading. Hence, the second step consists in multiplying each balance sheet

18 The haircut used for each asset class is the average across maturities. Calculations can be provided upon

request. See: https://www.ecb.europa.eu/ecb/legal/pdf/celex_32018o0004_en_txt.pdf

https://www.ecb.europa.eu/mopo/assets/risk/liquidity/html/index.en.html

29

category for the relative estimated coefficients to derive the numerator and denominator

of equation (17) for the sample of international banks and so obtaining the discount rate

(𝛿𝑖) and the pool of assets (𝜃𝑖).

Table 2: Step 1 - Regression Results for Euro Area Banks Sample

Note: Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

Figure 9 depicts respectively the bank-specific discount rate (𝛿𝑖) and the pool of

assets available for sale (𝜃𝑖), the latter as share of total assets. As can be noticed, the

bank-specific discount rate (𝛿𝑖) is centered around 57.5% and resembles a normal

distribution, whereas the pool of non-central bank eligible assets is left skewed, with a

mean centered around 4% of total assets and outliers reaching an amount higher than

20%.

Figure 9: Bank-Specific Discount Factor - Fire Sale Parameter

Source: COREP and FINREP Supervisory Data, Template F32.01 and bankscope.

Note: the 𝜹 coefficient reflects a weighted average haircut of the portfolio 𝜽 for non-central bank eligible

instruments.

EA Banks EA Banks

VARIABLES Coefficients Numerator Eq. 12 Denominator Eq. 12

Financial Assets Available for Sale (FAAS) (a1) 0.169*** 0.309***

-0.044 (0.0879)

Financial Assets Held for Trading (FAHT) (a2) 0.0641*** 0.108***

(0.00973) (0.0194)

High Quality Liquid Assets (HQLA) (a3) -0.0205 -0.0373

(0.0327) (0.0652)

Observations / 85 85

Adj.R-squared / 0.89 0.86

30

Overall, the nested set of liquidity and fire sale parameters (𝛾𝑖 , 𝜃𝑖 , 𝛿𝑖), depicted in

Figure 10, captures the degree of heterogeneity characterizing the liquidity strategies of

banks in our sample. For instance, a bank may choose to hold a larger amount of HQLAs

as a share of total assets (𝛾𝑖) - the area below the 45 degree line - than the pool of

unencumbered non-HQLA financial assets (𝜃𝑖) - the area above the 45 degree line. Banks

belonging to area (A) are those that may most likely suffer capital losses by liquidity

shocks since the liquidity surplus may easily become binding, and in turn may trigger fire

sales. On the contrary, banks belonging to area (B) are those that may most likely

experience a liquidity defaultwhen the liquidity surplus (𝛾𝑖) is depleted. In this case, the

pool of assets 𝜃𝑖 is likely to be insufficient to cover the remaining liquidity needs. In the

end, the quadrant (C) captures those banks that are short of both buffers and are clear

candidates for the realization of the liquidity default. Furthermore, the above mentioned

effects are far more pronounced when the size of the nodes are large (red nodes), since it

implies that they will face a harsher discount rate via fire sales. The realization of these

dynamics (A, B, C) is conditional to the amount of short-term bilateral

exposures 𝜌𝑖𝑥𝑖ℎ, which, in the end, determines the spread of contagion within the

interbank market.

Figure 10: Liquidity Default Dynamics

Note: nodes’ size is proportional to a bank’s discount rate (𝛿). Red nodes highlight the 75th percentile of the

discount factor. For confidentiality reasons, the chart shows statistics as average among three banks.

31

3.2.5 Distress-Default Threshold

A key assumption of the model is to define when counterparty bank i is not able to meet

its payment obligations, i.e. a default or distress threshold

(𝑐𝑖𝑑). Accordingly, a bank can be considered in default/distress when the surplus of

capital above the capital requirements a bank needs to meet at any time is depleted.

For our simulations we distinguish between two types of capital requirements: (i)

minimum capital requirements and (ii) capital buffers. The former requires banks to hold

4.5% RWAs of minimum capital (MC). This minimum requirement might be higher

depending on the bank-specific Pillar 2 requirement (P2R) set by the supervisor. In

addition to this, a bank is required to keep a capital conservation buffer (CCoB) of

between 1.875% and 2.5% CET1 capital as of 2018 (depending on the extent to which

the jurisdiction where the bank is located has fully or only partially phased in the end-

2019 requirement19

), and a bank-specific buffer, which is the higher among the Systemic

Risk Buffer (SRB), GSII and OSII buffers. Furthermore, some jurisdictions also apply a

positive counter-cyclical capital buffer requirement (CCyB). In this regard, we retrieved

bank-specific information on minimum capital requirements (CET1, TIER1, Own Funds)

and capital buffers from COREP supervisory templates C.01, C.03 and C.06.01 and the

bank-specific risk weighted assets (RWAs) from C.02. For international banks our data

source is Bankscope.

Therefore, the capital surplus (𝑘𝑖) can be defined in two ways: a capital surplus (𝑘𝑖𝐷𝐹)

above the minimum capital requirements defined as a default threshold (𝑐𝑖𝐷𝐹) reported in

equation (19a), or a capital surplus (𝑘𝑖𝐷𝑆) above the sum of the minimum capital

requirements and the capital buffers defined as a distress threshold (𝑐𝑖𝐷𝑆) presented in

equation (19b). When the bank breaches the minimum capital requirement (𝑐𝑖𝐷𝐹) it is

assumed that the supervisor would declare the bank for “failing or likely to fail” (which is

the official trigger for putting the bank into resolution).20

When the bank breaches the

19 In 2019 it will amount up to 2.5%. 20 As stated in the Bank Recovery and Resolution Directive (BBRD), the resolution authority should trigger

the resolution framework before a financial institution is balance sheet insolvent and before all equity has

been fully wiped out (Title IV, Chapter I, Art. 32, Point 41). Thus, our calibration method is consistent with

the Bank Recovery and Resolution Directive’s (BRRD) guidelines on fail or likely to fail: “An institution

shall be deemed to be failing or likely to fail in one or more of the following circumstances: … because the

32

buffer requirement (𝑐𝑖𝐷𝑆) while not yet breaching the minimum capital requirement, it is

assumed that it will not be declared failing but that it would rather be constrained in its

ability to pay out dividends. This, itself, could be a trigger for bank distress and is thus

considered as an alternative trigger threshold.

𝑘𝑖𝐷𝐹 = 𝑐𝑖 − 𝑐𝑖

𝐷𝐹

= 𝑐𝑖 − (𝑀𝐶𝑖 + 𝐶𝐶𝑜𝐵𝑖 + 𝑃2𝑅𝑖) (19𝑎)

𝑘𝑖𝐷𝑆 = 𝑐𝑖 − 𝑐𝑖

𝐷𝑆

= 𝑐𝑖 − [(𝑀𝐶𝑖 + 𝐶𝐶𝑜𝐵𝑖 + 𝑃2𝑅𝑖) + max (𝑆𝑅𝐵𝑖 , 𝐺𝑆𝐼𝐼𝑖 , 𝑂𝑆𝐼𝐼𝑖) + 𝐶𝐶𝑦𝐵𝑖] (19𝑏)

Hence, this calibration method allows for some flexibility on the determination of a

bank’s default depending on the purpose of the exercise. While from a resolution

authority and supervisory perspective the capital surplus (𝑘𝑖𝐷𝐹) based on the default

threshold (𝑐𝑖𝐷𝐹) may be the more relevant reference point, the distress threshold (𝑐𝑖

𝐷𝑆)

may be of interest to macroprudential supervisors. As this paper has a systemic risk

focus, we will provide results based on the latter approach. This is further motivated by

the fact that the inclusion of the macroprudential buffers (𝑆𝑅𝐵𝑖 , 𝐺𝑆𝐼𝐼𝑖 𝑂𝑆𝐼𝐼𝑖 , 𝐶𝐶𝑦𝐵𝑖)

allows to take into account the impact of macroprudential policy actions.21 Nevertheless,

we discuss the differences in the two approaches in the sensitivity analysis section.

Finally, an additional feature that needs to be taken into consideration in order to

accurately handle heterogeneity in the bank-specific capital surplus concerns the type of

capital used in the calculation. In fact, both the capital base (𝑐𝑖) and the minimum capital

(𝑀𝐶𝑖) and pillar 2 requirements (𝑃2𝑅𝑖) may vary whether the capital considered is

CET1, TIER1, or own funds calculated as the sum of TIER1 and TIER2 instruments. For

instance, 𝑀𝐶𝑖 are respectively 4.5% of RWAs for CET1 capital, 6% of RWAs for TIER1

capital, and 8% of RWAs for own funds. In turn, these differences are also reflected in

the capital base (𝑐𝑖). Hence, this implies that the very same bank may face a capital

surplus (𝑘𝑖𝐶𝐸𝑇1 ≷ 𝑘𝑖

𝑇𝐼𝐸𝑅1 ≷ 𝑘𝑖𝑂𝐹) larger or smaller depending on the capital considered.

institution has incurred or is likely to incur losses that will deplete all or a significant amount of its own

funds” (Title IV, Chapter I, Art. 32, Point 4). 21 Potentially also the Pillar 2 Guidance (P2G) may be included into the distress threshold calculation.

33

In this study, we use as benchmark the CET1 ratio, although we provide in the results

section evidence for the robustness of our findings to this calibration feature.

Figure 11: Bank-Specific Distress Threshold

Panel (a) Panel (b)

Source: COREP Supervisory Data Templates C.01-C.03, and Bankscope.

Note: The sum of minimum capital and capital surplus gives total capital (c). The decreasing ordering is

based on total capital. For confidentiality reasons, panel (b) shows surplus and minimum CET1 as average

among three banks following total capital decreasing ordering.

Panel (a) of Figure 11 depicts the distribution of the CET1 capital surplus based on a

distress threshold, while panel (b) presents the contribution of the capital surplus (distress

threshold) and the distress threshold to the capital base.22

Overall, the advantage of this methodology is twofold. It allows us to tailor a realistic

distress-default threshold and put it in relation to the bank’s voluntary buffer (here

defined as ‘capital surplus’) as well as to perform scenario analysis and counterfactual

analysis by imposing higher bank-specific capital requirements or by reducing the capital

surplus under an adverse scenario.

3.3 Model Outputs

This exercise is tailored to rank banks for their systemic risk contribution to financial

stability in terms of potential contagion and degree of vulnerability of the euro area

22 Although the average capital surplus varies little among the different capital classes (close to 7-8% of

RWAs). The number of banks close to the distress threshold moves from 24 for the CET1 capital threshold

to 22 for TIER1, and to 18 for Total Capital (Own Funds). Bank-specific distress threshold for Tier 1 and

total capital are available upon request.

34

banking system. Considering, a policy maker’s perspective, each bank is evaluated upon

four main model-based outputs, as follows:

i. Contagion index (CI): system-wide losses induced by bank i in percent of total

capital in the system (excluding bank i);

𝐶𝐼𝑖 = 100∑ 𝐿𝑗𝑖𝑗≠𝑖

∑ 𝑘𝑗𝑗≠𝑖 ,

where 𝐿𝑗𝑖 is the loss experienced by bank j due to the triggered default of bank i.

ii. Vulnerability index (VI): average loss experienced by bank i across all simulations

in percent of its own capital.

𝑉𝐼𝑖 = 100∑ 𝐿𝑖𝑗𝑗≠𝑖

∑ 𝑘𝑖𝑗≠𝑖 ,

iii. Contagion level: the number of banks that experience severe distress associated

with a default induced by the initial hypothetical failure of bank i;

iv. Vulnerability level: the total number of simulations under which bank i fails;

where 𝐿𝑖𝑗 is the loss experienced by bank i due to the triggered default of bank j.

Essentially, losses experienced by each bank (𝐿𝑗𝑖) is the sum of losses associated

with credit risk shock (𝐿𝐶𝑟𝑗𝑖) and losses associated with a funding risk shock (𝐿𝐹𝑢𝑗𝑖).

Hence, each index can be broken down to the respective contributions by credit risk

(CI_Cr and VI_Cr) and funding risk (CI_Fu and VI_Fu) shocks providing insights to the

nature of contagion.

𝐶𝐼_𝐶𝑟𝑖 = 100∑ 𝐿𝐶𝑟𝑗𝑖𝑗≠𝑖

∑ 𝑘𝑗𝑗≠𝑖 , 𝑎𝑛ⅆ 𝐶𝐼_𝐹𝑢𝑖 = 100

∑ 𝐿𝐹𝑢𝑗𝑖𝑗≠𝑖

∑ 𝑘𝑗𝑗≠𝑖 ,

𝑉𝐼_𝐶𝑟𝑖 = 100∑ 𝐿𝐶𝑟𝑖𝑗𝑗≠𝑖

∑ 𝑘𝑖𝑗≠𝑖 , 𝑎𝑛ⅆ 𝑉𝐼_𝐹𝑢𝑖 = 100

∑ 𝐿𝐹𝑢𝑖𝑗𝑗≠𝑖

∑ 𝑘𝑖𝑗≠𝑖 ,

The indices can be further decomposed based on banks’ geographical origins. For

example, CI_EAi is a sub-index based on the total induced losses by bank i to the subset

of banks that are in the euro area. Similarly, VI_EAi is a sub-index based on the average

losses experienced by bank i across the subset of simulations where the triggered banks

are in the euro area. Essentially, these two indices capture a given bank’s contagion and

vulnerability to euro area banks.

35

𝐶𝐼_𝐸𝐴𝑖 = 100∑ 𝐿𝑗𝑖𝑗≠𝑖

∑ 𝑘𝑗𝑗≠𝑖 , 𝑖 ∈ 𝕊𝐸𝐴 , 𝑎𝑛ⅆ 𝑉𝐼_𝐸𝐴𝑖 = 100

∑ 𝐿𝑖𝑗𝑗≠𝑖

∑ 𝑘𝑖𝑗≠𝑖 , 𝑗 ∈ 𝕊𝐸𝐴

where 𝕊𝐸𝐴 is the subsample of banks in the euro area.

The geographical focus can be based on distinguishing between euro area and non-

euro area banks as well as between individual countries where banks are domesticated.

Moreover, based on these outputs, we develop two additional indicators to deepen both

analytical assessment and policy implications of contagion analysis and in turn facilitate

the impact assessment of regulatory actions.

v. Amplification ratio (cascade effect): this metric compares the losses induced by a

bank’s simulated default in the initial round vis-à-vis those occurring in all successive

rounds. From the perspective of a bank, bank i, triggering system-wide contagion:

𝐴𝑀𝑃(𝐶)𝑖 =∑ 𝐿𝑗𝑖 𝑟0+𝑡𝑗≠𝑖

∑ 𝐿𝑗𝑖 𝑟0𝑗≠𝑖

𝑊ℎ𝑒𝑟𝑒 𝑟0 = 𝑖𝑛𝑖𝑡𝑖𝑎𝑙 𝑟𝑜𝑢𝑛ⅆ ; 𝑟0+𝑡 = 𝑠𝑢𝑐𝑐𝑒𝑠𝑠𝑖𝑣𝑒 𝑟𝑜𝑢𝑛ⅆ𝑠.

This index measures how much of the system-wide impact from the failure of bank i is

caused by cascading of defaults rather than direct and immediate losses from bank i.

Hence, the higher the ratio, the larger the amplification through the network, and a ratio

greater than 1 indicates that losses due to cascade effects dominate direct losses.

Conversely, banks’ susceptibility to systemic events can be split into two similar

components to distinguish how much of the losses experienced by bank i across all

simulations were immediate losses as opposed to losses in successive rounds:

𝐴𝑀𝑃(𝑉)𝑖 =∑ 𝐿𝑖𝑗 𝑟0+𝑡𝑗≠𝑖

∑ 𝐿𝑖𝑗 𝑟0𝑗≠𝑖

Amplification effect quantifies the degree to which cascading behavior impacts the

banks both at system-wide and entity level. These are the losses associated with

contagion spread through indirect linkages. Amplification effect is an important metric

of the financial system architecture in the sense that it captures what portion of systemic

risk is not directly observable to banks and, possibly, to the regulators in the absence of

granular data.

36

vi. Sacrifice ratio: One of the most important policy questions during financial crises or

near-crisis episodes concerns bank recapitalization or emergency liquidity assistance.

The concept of “too big to fail” does not solely depend on a bank’s size and portfolio

compared to the domestic system but also on whether a single failure would induce a

larger collapse in the financial system. With respect to the latter, we construct a

sacrifice ratio, which measures the ratio of system-wide losses due to the failure of a

bank over the cost of a rescue package, tax-payer sacrifice, equal to the capital

requirements of the bank.23

𝑆𝑅𝑖 =∑ 𝐿𝑗𝑖𝑗≠𝑖

𝑐𝑖𝐷𝑆

Therefore, ratios above 1 are associated with system-wide losses that could be

avoided with relatively smaller cost to the tax payer. In contrast, ratios smaller than 1

would imply that potential system-wide losses are not sufficiently large to justify the

government assistance to the bank. We provide three types of sacrifice ratios from the

perspectives of: (i) a global central planner; (ii) a euro area authority; and (iii) a national

authority. These three measures take into account system-wide losses respectively

induced to all banks in the system (global central planner) or to those banks belonging

to each jurisdiction, whether it is euro area based, or national. It is important to

underline that our modelling strategy does not take into account the mitigation and

amplification effects induced by a bail-in mechanism, neither for the triggering bank nor

for the subsequent failing banks.24

Once the magnitude of system-wide losses associated

with bank failures are understood, a critical policy question is how the regulators

respond to the pending default of an entity. In other words, the regulators would need to

know whether the public cost of making the entity whole again justifies the potential

damages to the system when no action is taken.

23 Capital requirements are defined as for the distress threshold as follows: (𝑀𝐶𝑖 + 𝐶𝐶𝑜𝐵𝑖 + 𝑃2𝑅𝑖) +max(𝑆𝑅𝐵𝑖 , 𝐺𝑆𝐼𝐼𝑖 , 𝑂𝑆𝐼𝐼𝑖) + 𝐶𝐶𝑦𝐵𝑖 . 24 Introducing a bail-in mechanism into the model would tend to reduce the cost of a bank-recapitalization

since bailinable liabilities would be transformed into new equity of the bank considered for resolution thus

shielding taxpayers. At the same time, it could increase the amount of losses experienced by the other

banks in the network since creditors’ assets such as additional tier 1 and tier 2 instruments, other

subordinated debts, senior unsecured debt and non-eligible deposits, and non-covered eligible deposits may

face a partial or complete written-off (see Hüser et al., 2017).

37

4 Results

4.1 Main Findings

This subsection delves in greater detail into main findings of the exercise across a broad

range of interconnectedness attributes based upon our benchmark model with bank-

specific calibration. For the sake of clarity, out of 199 worldwide consolidated banking

groups, Table 3 reports the top-50 default events ranked in terms of contagion index (CI)

to the euro area banking system.

The scale of losses follows a power-law distribution, with the top-10 banks inducing

on average 2.5% of capital losses to the euro area banking system, around EUR 25

billion.25

The CI index of the most contagious bank is larger by a factor of 12 than the

average of the full sample, and even among the top-10 most contagious banks remarkable

differences exist. International spillovers seem to be a key channel of contagion to the

euro area banking system, with 27 banking groups out of the top-50 located outside the

euro area. Therefore cross-border risks may propagate quickly via bilateral exposures to

the euro area banking network, as evident during the Great Financial Crisis of 2008.

In terms of channels underlying contagion, losses due to credit risk dominate those

due to funding/liquidity risk via fire sales. When it comes to the nature of defaults

(contagion level), illiquidity-driven defaults outweigh by far those triggered due to

insolvency matching the historical reality as emphasized by Aikman et al. (2018).

Notably, solvency defaults are mostly concentrated in the top-10, underlying how

contagion due to solvency risks may be highly concentrated on few players due to their

central role in the network as borrowers; whereas liquidity risks may be triggered by a

greater number of lenders in the network for which the intrinsic characteristics of the

borrower play a more crucial role. In other words, funding risk relative to credit risk

seems to be less diversifiable and more concentrated on few large exposures, whose

defaults may trigger a liquidity default event. This finding may suggest that it may be

prudent to impose limits on the concentration of funding since some small and medium-

sized banks are exposed on the liability side to few large banks.

25 Result for power-law distribution can be provided upon request. The capital of the triggering bank is not

included in the CI index calculation.

38

Table 3: Contagion Measures

Note: For confidentiality reasons bank names have been anonymized. The results in this table are ranked by

CI index. CI refers to contagion index at euro area scale and amounts represent capital losses to all banks in

percent of entire banking system’s total capital. This index is further decomposed into the respective

contributions by credit (CI CR) and funding (CI FU) shocks. Defaults refer to the number of defaults a

bank has induced in the system. Rounds indicate the maximum number of rounds the simulation required

until no additional defaults in the system, whereas amplification ratio is the ratio of losses in subsequent

rounds to losses in the initial round. The sacrifice ratio indicates the ratio of systemic losses caused by a

bank over the cost of rescue package to fully recapitalize the bank.

RANK Country CI CI CR CI FU ToT Solv. Liq. Rounds Ratio G EA N

1 EA 3.8 3.8 0.0 0 0 0 1 0.0 0.6 0.6 0.2

2 XEA 3.2 3.1 0.1 3 1 2 2 0.1 0.4 0.4 0.0

3 XEA 2.7 2.3 0.4 1 1 0 2 0.5 0.6 0.4 0.1

4 EA 2.4 2.4 0.0 2 2 0 2 0.0 0.7 0.7 0.3

5 EA 2.3 2.3 0.0 0 0 0 1 0.0 0.7 0.7 0.3

6 XEA 2.3 2.1 0.2 3 0 3 2 0.2 0.9 0.7 0.1

7 EA 2.2 2.1 0.1 1 0 1 2 0.0 0.4 0.4 0.2

8 XEA 1.9 1.9 0.0 0 0 0 1 0.0 0.3 0.3 0.0

9 XEA 1.9 1.9 0.0 0 0 0 1 0.0 0.5 0.5 0.0

10 EA 1.9 1.9 0.0 0 0 0 1 0.0 0.5 0.4 0.3

AVERAGE TOP 10 2.5 2.39 0.07 1.0 0.4 0.6 1.50 0.1 0.56 0.52 0.13

11 EA 1.8 1.8 0.0 1 1 0 2 0.0 0.3 0.3 0.1

12 XEA 1.4 1.2 0.2 1 0 1 2 0.1 0.3 0.2 0.0

13 XEA 1.4 1.3 0.1 2 0 2 2 0.3 0.9 0.7 0.0

14 EA 1.3 1.3 0.0 0 0 0 1 0.0 0.4 0.4 0.0

15 XEA 1.1 1.0 0.1 0 0 0 1 0.0 2.2 2.2 0.0

16 XEA 1.0 1.0 0.0 0 0 0 1 0.0 0.6 0.6 0.0

17 EA 1.0 1.0 0.0 0 0 0 1 0.0 0.5 0.4 0.2

18 EA 0.9 0.9 0.0 0 0 0 1 0.0 0.3 0.3 0.1

19 EA 0.8 0.8 0.0 1 0 1 2 0.1 0.4 0.4 0.3

20 XEA 0.8 0.8 0.0 0 0 0 1 0.0 0.3 0.3 0.0

21 XEA 0.8 0.8 0.0 0 0 0 1 0.0 0.5 0.5 0.0

22 EA 0.8 0.4 0.4 1 1 0 2 4.3 5.0 1.8 0.8

23 XEA 0.8 0.8 0.0 0 0 0 1 0.0 0.9 0.9 0.0

24 EA 0.8 0.8 0.0 0 0 0 1 0.0 0.5 0.5 0.2

25 XEA 0.7 0.7 0.0 0 0 0 1 0.0 0.3 0.3 0.0

26 EA 0.6 0.6 0.0 0 0 0 1 0.0 0.2 0.2 0.0

27 XEA 0.6 0.6 0.0 0 0 0 1 0.0 0.0 0.0 0.0

28 XEA 0.6 0.6 0.0 0 0 0 1 0.0 0.5 0.5 0.0

29 EA 0.5 0.5 0.0 0 0 0 1 0.0 1.2 1.2 1.2

30 EA 0.5 0.4 0.1 0 0 0 1 0.0 11.4 11.4 7.2

31 EA 0.5 0.5 0.0 0 0 0 1 0.0 1.2 1.1 1.1

32 XEA 0.5 0.5 0.0 0 0 0 1 0.0 0.1 0.1 0.0

33 XEA 0.5 0.5 0.0 0 0 0 1 0.0 0.0 0.0 0.0

34 XEA 0.5 0.4 0.1 3 0 3 2 1.4 1.3 0.5 0.0

35 EA 0.5 0.5 0.0 0 0 0 1 0.0 0.3 0.3 0.2

36 EA 0.4 0.4 0.0 0 0 0 1 0.0 0.2 0.2 0.0

37 XEA 0.4 0.4 0.0 0 0 0 1 0.0 0.3 0.3 0.0

38 XEA 0.4 0.4 0.0 0 0 0 1 0.0 0.0 0.0 0.0

39 EA 0.4 0.4 0.0 0 0 0 1 0.0 5.8 1.3 0.1

40 XEA 0.4 0.4 0.0 0 0 0 1 0.0 0.0 0.0 0.0

41 XEA 0.4 0.4 0.0 0 0 0 1 0.0 1.4 1.4 0.0

42 EA 0.4 0.4 0.0 0 0 0 1 0.0 4.0 3.9 1.1

43 XEA 0.4 0.4 0.0 0 0 0 1 0.0 1.2 1.2 0.0

44 XEA 0.4 0.4 0.0 0 0 0 1 0.0 0.3 0.3 0.0

45 XEA 0.4 0.4 0.0 0 0 0 1 0.0 0.1 0.1 0.0

46 EA 0.4 0.4 0.0 0 0 0 1 0.0 2.4 0.3 0.1

47 XEA 0.3 0.3 0.0 0 0 0 1 0.0 0.7 0.7 0.0

48 XEA 0.3 0.3 0.0 0 0 0 1 0.0 0.0 0.0 0.0

49 EA 0.3 0.3 0.0 0 0 0 1 0.0 0.5 0.5 0.5

50 EA 0.3 0.3 0.0 0 0 0 1 0.0 0.4 0.3 0.3

AVERAGE TOP 50 1.0 0.98 0.03 0.38 0.12 0.26 1.22 0.14 1.05 0.82 0.30

0.3 0.31 0.01 0.10 0.03 0.07 1.06 0.04 0.77 0.67 0.15

Contagion

AVERAGE

EURO AREA Sacrifice RatioDefaults Amplification

39

4.1.1 Amplification Effects

One powerful feature of our framework is its ability to capture cascade effects due to an

initial distress event. In most cases, the contagion does not spread beyond the direct

counterparties in the first round. At most, the contagion cycle ends after a second round

of failures, and overall has limited repercussions on the system as a whole. There are only

two banks with an amplification effect larger than 1 and whose failure causes losses 4.3

and 1.4 times more in subsequent rounds compared to the initial round.

4.1.2 Sacrifice ratio

In this exercise, four failures warrant an intervention in the form of a rescue package

by national authorities. The average sacrifice ratio is close to 0.3 for the top-50, whereas

for the top-10 most systemic banks is equal to 0.13. The latter is lower than the former

because i) global banks tend to have most of their exposures outside the domestic

interbank market, thereby implying a lower numerator of the sacrifice ratio than a bank

more domestic-oriented, ii) global banks tend to have a larger denominator than the more

domestic-oriented ones due to higher capital requirements. However, if supervisors take

the perspective of a euro area authority, the average ratio of the top-50 increases up to

0.82, and five more interventions are justified. In this respect, six out of nine defaults

originate from entities located within the euro area and, thus, can be resolved within the

realm of the Single Resolution Mechanism. The other three interventions should be

granted to extra-euro area banks, and given their respective contagion to the euro area

banking system, it might be instructive to monitor the evolution of such costly spillovers

and cooperate closely with international authorities as needed.

In the end, a global perspective increases the number of interventions up to eleven,

thereby highlighting how international cooperation is necessary to reduce negative

externalities to the whole financial system

4.1.3 Vulnerability

Having investigated the various aspects of contagion, it is important to understand its

complementary interface, i.e. which banks are the most vulnerable and how contagion

affects them. Hence, banks are ranked by vulnerability index at global scale.

40

Table 4: Vulnerability Measures

Note: For confidentiality reasons bank names have been anonymized. VI refers to vulnerability index and

amounts represent average capital losses across all independent simulations in percent of a bank’s capital.

The vulnerability index is further decomposed into the respective contributions by credit (VI CR) and

funding (VI FU) shocks. VI from Euro Area (VI EA) is computed with respect to average losses caused by

banks in respective groups. Defaults refer to the number of defaults a bank has experienced given the

hypothetical (exogenous) defaults of each other bank in the system. Amplification ratio is is the ratio of

losses in subsequent rounds to losses in the immediate round. The results in this table are ranked by VI

Global Scale.

Amplification Contribution

RANK Country VI VI CR VI FU ToT Solv. Liq. Ratio Ratio VI VI CR VI FU

1 EA 1.8 0.7 1.2 2 2 0 0.0 0.9 1.6 0.7 1.0

2 EA 1.4 1.4 0.0 0 0 0 0.0 0.7 1.0 1.0 0.0

3 EA 1.1 1.1 0.0 0 0 0 0.0 1.5 1.6 1.6 0.0

4 EA 1.0 1.0 0.0 0 0 0 0.0 0.8 0.8 0.8 0.0

5 EA 0.9 0.9 0.0 0 0 0 0.0 1.1 1.0 1.0 0.0

6 XEA 0.8 0.8 0.0 0 0 0 0.4 1.1 0.9 0.9 0.0

7 EA 0.8 0.8 0.0 0 0 0 0.0 0.6 0.4 0.4 0.0

8 EA 0.7 0.7 0.0 0 0 0 0.0 0.9 0.6 0.6 0.0

9 EA 0.6 0.6 0.0 0 0 0 0.0 0.4 0.3 0.3 0.0

10 EA 0.6 0.5 0.1 2 0 2 0.0 2.1 1.3 1.1 0.2

AVERAGE TOP 10 1.0 0.8 0.1 0.4 0.2 0.2 0.0 1.0 0.9 0.8 0.1

11 EA 0.6 0.6 0.0 0 0 0 0.2 1.4 0.9 0.9 0.0

12 EA 0.5 0.5 0.0 0 0 0 0.0 0.7 0.4 0.4 0.0

13 EA 0.5 0.5 0.0 0 0 0 0.0 0.6 0.3 0.3 0.0

14 EA 0.5 0.5 0.0 0 0 0 0.0 0.8 0.4 0.4 0.0

15 EA 0.5 0.5 0.0 0 0 0 0.1 1.2 0.6 0.6 0.0

16 EA 0.4 0.4 0.0 0 0 0 0.1 1.0 0.4 0.4 0.0

17 EA 0.4 0.4 0.0 0 0 0 0.0 1.2 0.5 0.5 0.0

18 EA 0.4 0.4 0.0 3 3 0 0.0 1.3 0.5 0.5 0.0

19 EA 0.4 0.4 0.0 0 0 0 0.0 1.8 0.8 0.8 0.0

20 EA 0.4 0.4 0.0 0 0 0 0.0 1.0 0.4 0.4 0.0

21 EA 0.4 0.4 0.0 0 0 0 0.0 0.7 0.3 0.3 0.0

22 EA 0.4 0.4 0.0 0 0 0 0.0 2.1 0.8 0.8 0.0

23 EA 0.4 0.4 0.0 0 0 0 0.0 0.8 0.3 0.3 0.0

24 EA 0.4 0.4 0.0 0 0 0 0.0 1.0 0.4 0.4 0.0

25 EA 0.4 0.4 0.0 0 0 0 0.0 2.1 0.7 0.7 0.0

26 EA 0.4 0.4 0.0 0 0 0 0.0 1.6 0.6 0.6 0.0

27 EA 0.3 0.3 0.0 0 0 0 0.0 0.9 0.3 0.3 0.0

28 EA 0.3 0.3 0.0 0 0 0 0.0 1.0 0.4 0.4 0.0

29 EA 0.3 0.3 0.0 0 0 0 0.0 1.2 0.3 0.3 0.0

30 EA 0.3 0.3 0.0 0 0 0 0.0 0.6 0.2 0.2 0.0

31 EA 0.3 0.3 0.0 0 0 0 0.0 1.8 0.5 0.5 0.0

32 EA 0.3 0.2 0.1 4 0 4 0.0 0.1 0.0 0.0 0.0

33 EA 0.3 0.3 0.0 0 0 0 0.0 0.7 0.2 0.2 0.0

34 EA 0.3 0.1 0.1 4 0 4 0.0 0.2 0.0 0.0 0.0

35 EA 0.3 0.3 0.0 0 0 0 0.0 0.9 0.2 0.2 0.0

36 EA 0.3 0.3 0.0 0 0 0 0.0 0.8 0.2 0.2 0.0

37 EA 0.2 0.2 0.0 0 0 0 0.0 2.1 0.5 0.5 0.0

38 XEA 0.2 0.2 0.0 0 0 0 0.9 1.3 0.3 0.3 0.0

39 EA 0.2 0.2 0.0 0 0 0 0.0 1.8 0.4 0.4 0.0

40 EA 0.2 0.2 0.0 0 0 0 0.0 0.0 0.0 0.0 0.0

41 EA 0.2 0.2 0.0 0 0 0 0.0 1.1 0.2 0.2 0.0

42 EA 0.2 0.2 0.0 0 0 0 0.0 2.1 0.4 0.4 0.0

43 EA 0.2 0.2 0.0 0 0 0 0.0 0.8 0.2 0.2 0.0

44 EA 0.2 0.2 0.0 0 0 0 0.0 1.9 0.4 0.4 0.0

45 EA 0.2 0.2 0.0 0 0 0 0.0 1.3 0.3 0.3 0.0

46 EA 0.2 0.2 0.0 0 0 0 0.0 0.8 0.2 0.2 0.0

47 EA 0.2 0.2 0.0 0 0 0 0.0 1.2 0.2 0.2 0.0

48 EA 0.2 0.2 0.0 0 0 0 0.0 0.8 0.2 0.2 0.0

49 EA 0.2 0.2 0.0 0 0 0 0.0 1.3 0.3 0.3 0.0

50 EA 0.2 0.2 0.0 0 0 0 0.0 0.6 0.1 0.1 0.0

0.45 0.42 0.03 0.3 0.1 0.2 0.03 1.1 0.5 0.4 0.0

0.14 0.13 0.01 0.1 0.0 0.1 0.05 1.3 0.2 0.1 0.0AVERAGE

GLOBAL SCALE EURO AREADefaultsVulnerability

AVERAGE TOP 50

41

Focusing on the banks with the 50 highest vulnerability scores, they are almost all

from within the euro area. This is due to the fact that the large exposures dataset, as

emphasized in section 2, mostly captures exposures from euro area banks, and for this

precise reason, we adopt a euro area centric view. Although, this index is computed as the

share of losses experienced by a bank over the bank’s capital base, and not in % of the

system capital, the distribution of losses also for this index follows a power-law

distribution.

As previously emphasized, funding risk is concentrated with few banks, whereas

losses due to credit risks are spread over the entire sample. Consistently, out of 15

defaults experienced by the top-50 most vulnerable banks, liquidity defaults accounts for

2/3 of failures, while solvency defaults account only for 1/3. Cascade effects seem to play

a minor role in terms of loss amplification and induced defaults. None of the banks

defaulting shows a positive amplification ratio, which is concentrated to a few entities.

We also report for comparative purpose, the euro area based vulnerability index and

its regional contribution in order to disentangle which banks may be most vulnerable

from shocks arising from within and outside the euro area banking system. On average,

losses produced from within the euro area are 1.3 time higher than the amount of losses

experienced from entities located outside the euro area. Whereas, for the top-10 most

vulnerable banks externally-driven losses are at least as important as euro area induced

losses, reflecting the international profile of this group of institutions.

In Figure 12, a systemic risk map combines information from both indexes in order

to allow an easy identification of threats to euro area financial stability.26

In this picture,

the contagion and vulnerability indexes are normalized and reported in absolute terms,

i.e. total amount of losses induced and experienced by each bank, respectively. Figure 12

is divided into four quadrants capturing different degrees of banks’ systemic footprints.

Banks in the north-west quadrant (B) are those whose default would induce the greater

amount of losses to the euro area banking system, while those lying in the south-east

quadrant (C) are those most vulnerable to a default event. Banks located in the north-east

26 The vulnerability index is here reported in absolute terms, i.e. considering the EUR value of capital

depletion, and not as a % of the capital base.

42

quadrant (D) are both highly contagious and vulnerable. These metrics provide a useful

monitoring tool to assess vulnerabilities in the euro area banking system due to

interconnectedness.

Figure 12: Systemic-Risk Map

Note: Contagion and vulnerability indexes are normalized by dividing each index for the entity’s maximum

value.

4.2 Bank-Specific Calibration

Parameters and model heterogeneity renders a different picture compared with

homogeneous parameters ( 𝜆, 𝜌, 𝛾, 𝜃, 𝛿 ) approximated as sample averages uniformly

applied to all banks (Special Case I). This difference is even more marked when we don’t

consider the bank-specific liquidity coverage ratio (𝛾) and the liquidity default

assumptions, respectively by setting 𝛾 = 0 and 𝜃 = ∞ (Special Case II).27

As shown in

Figure 13, in both cases losses are larger than in the bank-specific calibration (benchmark

case). On the one hand - special case I - using average parameters cancels all the losses

coming from funding risk (See Table 1A Appendix). This is due to the fact that the

liquidity buffer (𝛾) is large enough to cover all short-term funding needs. On the other

hand, a bank-average calibration neglecting the liquidity buffer (𝛾) and the constraint

based on the pool of available for sale assets (𝜃) results in an over-estimation of funding

27 For computing fire sales losses, we now use a lower discount rate than the average of the bank-specific

one because assets available for sales are not limited to the pool of non-HQLA assets, but comprehend also

HQLA assets which, by definition, face lower haircuts. The former average was set close to 57.5% (as

shown in figure 9), while now it stands to 26%.

43

risk. This is due to the assumption that each funding shortfall is directly transformed via

fire sales into a solvency risk.

Figure 13: Bank-Specific Calibration

Note: Benchmark refers to results presented in section 4.1. Special case I sets ( 𝜆, 𝜌, 𝛾, 𝜃, 𝛿 ) equal to the

average of the sample, while special case II sets ( 𝜆, 𝜌, 𝛿 ) equal to the average of the sample and 𝛾 = 0 and

𝜃 = ∞. We use the same distress threshold as in the benchmark case.

These findings highlight that bank-heterogeneity is an essential determinant of

liquidity contagion, since weak nodes are those channels amplifying the initial shock and

creating cascade effects. Moreover, by applying average parameters credit risk increases.

The intuition behind this effect is that banks with larger and riskier exposures tend to ask

for a higher collateral amount, which results in a counterbalancing-risk behavior and a

lower exposure-specific LGD vis-à-vis riskier counterparties.

Overall, the results from this exercise support the importance of calibrating the model

with bank-level specificities and how the inclusion of prudential regulations into the

model parameters such as the HQLA buffer (𝜆) is essential to more properly estimate

liquidity risk. Model and parameter heterogeneity lead to significant corrections and

overall changes in the CI and VI indexes as well as in bank-specific scores. In addition,

this exercise clearly highlights the role played by weak nodes in amplifying contagion,

and how neglecting them may lead to estimation bias.

44

4.3 Macro Stress Test Scenarios

The framework has so far been applied as an exercise that simulated the hypothetical

failure of each bank separately. Bank defaults may occur for purely institution-specific

reasons. However, often bank defaults (or distress) happen against the background of

more widespread stress in the financial system. Under such stressed circumstances the

contagion potential from one bank defaulting might be more pronounced as also the

banks’ counterparts are in a weakened position. To explore the contagion risk under a

generalized financial stressed situation, we incorporate the EBA 2016 stress test adverse

scenario in our framework.28

In this exercise, as shown in equation (20), we recalibrate the capital surplus

(𝑘𝑖𝐷𝑆_𝑆𝑇16

) in line with the capital depletion (𝑐𝑖𝑆𝑇16) that resulted from the 2016 EBA

Stress Test of EU banks under an adverse scenario (Figure 14).29

𝑘𝑖𝐷𝑆_𝑆𝑇16 = 𝑘𝑖

𝐷𝑆 − 𝑐𝑖𝑆𝑇16 (20)

Figure 14: Distress Threshold – Stress Test 2016’s Capital Depletion

Source: COREP Supervisory Data Templates C.01 – C.03, and Bankscope.

Note: The sum of minimum CET1, surplus CET1 and ST16 gives total CET1. ST16 refers to the capital

depletion due to the 2016 stress test’s results. For confidentiality reasons, the chart shows surplus and

minimum CET1 as average among three banks following total capital decreasing ordering.

Overall, as shown in Figure 15, the weakened solvency position of euro area banks

results in a disproportional increase of the contagion index. The top-10 most systemic

28 As an alternative to assuming a macro shock, we could implement a multiple default scenario

approximating a severe shock causing near collapse of the banking system in several major economies (i.e.

Global Financial Crisis). 29 We apply the average capital depletion to those banks not included in the EBA exercise.

45

banks increase their average losses caused at aggregate level by 1.25 times. In addition,

accounting for an adverse macro scenario reshapes the ranking of the most systemic

banks. In contrast, the vulnerability ranking is more affected at the center of the

distribution, with the most vulnerable banks preserving their position. Also concerning

this indicator, the effects seem to suggest non-linearities as shown by uneven changes

across banks. Overall, the average number of induced defaults rises by a factor of 3.3

(from 0.1 to 0.34), impacting consistently both the average top-50 amplification ratio and

the sacrifice ratio, which respectively increase from 0.05 to 0.25 and from 0.52 to 0.61

(for the euro area authority), with 2 additional cases of positive system-wide returns from

banks’ recapitalization.

Figure 15: Non-linear Effects of a Stress Test Scenario

Note: Benchmark refers to results presented in section 4.1.

4.4 Sensitivity Analysis

In this section we test the sensitivity of the results to a range of parameter assumptions so

as to disentangle the key determinants of contagion and vulnerability of the euro area

banking system.

The first sensitivity analysis tests the difference between a distress and a default

threshold as described respectively in equation (19a and 19b). As can be observed in

Figure 16 (Appendix), on average the affected banks operate with a capital surplus based

on a default threshold which is 1% higher than the capital surplus based on a distress

threshold (in RWA terms), with some outliers close to 3% (panel a). Nonetheless, as we

46

can see from panel (b) this difference does not produce any relevant variation in the

contagion and vulnerability indexes. The explanation for this finding is that the most

systemic banks are also the ones facing the highest macroprudential buffer

requirements (𝑆𝑅𝐵𝑖 , 𝐺𝑆𝐼𝐼𝑖 , 𝑂𝑆𝐼𝐼𝑖) and generally did not fail in the benchmark exercise

with a distress threshold, which is by construction smaller than the default threshold.

Moreover, the CCyB buffer is currently of a small magnitude, thereby reducing the

capital surplus of the distress threshold by little.

A second robustness check aims to verify whether changing the capital base from

CET1 to own funds (total capital) may affect the capital surplus based on the distress

threshold in a sizeable manner. As reported in Figure 17 panel (a) in the appendix, banks

may face both a decrease or an increase in the capital surplus depending on whether the

increased amount of minimum capital requirements (from 4.5% CET1 to 8% Own Funds)

outweighs the increased amount of capital included into own funds, i.e. additional tier 1

and tier 2 instruments. Overall, the results seem to be almost invariant to the selection of

the capital base. In both exercises, the findings remain unchanged relative to the

benchmark case.

The third exercise aims to test the sensitivity of the results to a deterioration of

banks’ liquidity position. In this respect, the key liquidity parameters, respectively the

funding shortfall (𝜌𝑖), the net liquidity position (𝛾𝑖) and the pool of assets available for

sales (𝜃𝑖) are adjusted to intensify a liquidity shock conditional on a default event. As we

can see from panel (a) of Figure 18 (appendix) an increase of the short-term funding from

an average of 35% (one month in the benchmark case) to 40% (3 months average) or

alternatively an increase of short-term funding up to 50% produces a negligible effect on

the contagion index, with only two banks on the contagion side facing a relevant positive

adjustment.30

This implies that the current LCR ratio (𝛾𝑖) is large enough to cover the

additional short-term liquidity needs. Next, we suppose a reduction of 20% and 40% in

the net liquidity position (𝛾𝑖) for instance due to a sudden depletion of the buffer of

HQLAs or due to a higher run off rate of deposits (panel b). In this case, we notice that

results are almost unaffected for both indexes, implying that the bank-specific liquidity

30 The average funding shortfall increases from 35% to 40% when considered 3 months as short-term

liabilities.

47

buffer is well above the threshold needed to cover short-term liquidity needs. Finally, we

assume that the pool of assets available for sales faces a haircut of 20% and 40%

respectively (panel c). Results are unchanged because no other banks, except those

previously triggering the fire sales stage, are liquidity constrained and therefore no

additional losses or failures are accounted in the system given this assumption.31

The fourth sensitivity analysis aims at capturing the role played by the network

structure. In achieving this, we exploit additional granular exposure-level information

from COREP template C.28 regarding bilateral linkages. Hence we test the sensitivity of

the contagion and vulnerability indexes to a network structure based on total gross

amounts (including exemptions). As we can see in Figure 19, when we consider the

network structure based on gross amounts (€ 1.120 bn) both contagion and vulnerability

indexes, as expected, strongly increase. These effects seem to be strong and distributed

unevenly pointing to non-linear effects. Several banks move from the bottom to the top of

the contagion index, depleting almost 3-4% of the capital of the system given their

default. Regarding the vulnerability index, the effects are even more pronounced than the

contagion index, and the largest increase is about a factor of 55, depleting on average

33% of the capital base.

The fifth exercise consists of testing the sensitivity of the results to the interaction of

different dimensions as in Kok and Montagna (2016): liquidity, solvency and network

topology. In this respect, we assume first a full-liquidity shock affecting all parameters as

in the benchmark case of the third sensitivity analysis, but contemporaneously: (𝜌𝑖 =

40%; ∆𝛾𝑖 = −20%; ∆𝜃𝑖 = −20%). Then we decrease by 20% the capital surplus, and we

proportionally increase by 30% the exposure amounts which correspond to the gross

amount presented in the fourth sensitivity analysis (€ 1.120 billion). As we can see from

Figure 20 (appendix) the interplay of liquidity parameters pushes up the contagion and

vulnerability indexes of some specific banks, but does not result in a relevant system-

wide increase. When this effect is combined with a decreased capital surplus, more

indices increase, but also in this case it does not produce a systemic-wide effect. When

31 The vulnerability index decreases for some selected banks because now those banks can sell a less

amount of assets and therefore will face a lower amount of losses via fire sales. Nevertheless, the number

of liquidity defaults remains unchanged.

48

we also stretch the network structure we are able to capture, this time, a quasi-linear

effect on contagion and vulnerability indexes. The few non-linear effects are produced by

the interaction of liquidity and solvency parameters. This exercise once more highlights

the role played by tipping points within a network and how bank-specific and exposure-

specific characteristics may determine non-linear amplification effects resulting in a

higher degree of systemic-risk.

Finally, we compare our model-based estimates of contagion and vulnerability

indices to market data-based measures of individual banks’ systemic footprint. We use

the SRISK index based on Global Dynamic Marginal Expected Shortfall (MES) retrieved

from V-lab and based on Acharya et al. (2012).32

For this purpose, we download and

adjust the SRISK index for the European Union such that it overlaps with our sample of

euro area banks. By doing this, we end up with 40 banks.33

Then we construct a SRISK

index based on our model-based estimates by dividing the numerator of the vulnerability

index (total losses experienced by each bank) by the total losses experienced by the

system. By doing this, we are able to derive the proportional contribution of each firm's

SRISK to the total positive SRISK of the euro area banking system. Figure 21 depicts

both SRISK index and the ranking based on the SRISK index. Both measures display a

high correlation with VI of 0.85 for the SRISK index and 0.74 for the SRISK ranking,

respectively. As can be seen in the bottom left of the SRISK ranking, for the top-10 banks

our balance-sheet based approach captures the very same banks of the SRISK approach

based on MES.

Overall, we find both similarities and divergences between market-based measures

such as SRISK and balance sheet based measures such as our approach. Notwithstanding

the correspondence (high correlation) between our VI measure and SRISK is reassuring,

we identify notable differences for individual banks in the middle of the distribution. This

is potentially due to our modelling strategy which takes into account bank-specific

characteristics which are only implicitly reflected in the market-based measures such as

SRISK.

32 See also Acharya et al. (2017). 33 Differences in the sample are due to the fact that Acharya et al. (2010) work on a solo basis and they

don’t include banks for which market-data are not available. The SRISK estimates are based on a 40% fall

of the broad market index and a 8% capital requirement.

49

5. Reducing contagion risk by fine-tuning prudential measures

The modelling framework can also be used to conduct ex ante impact analysis of

prudential policy measures based on counterfactual analyses. By exploiting the

breakdown of the vulnerability index into credit and funding risks, we are able to target

banks with specific liquidity and capital vulnerabilities that may give rise to contagion.

Our findings suggest that regulators should look at the interplay of network topology

and bank-specific characteristics in addition to setting prudential requirements based on

individual supervisory assessments of each bank in isolation. Tipping points shifting the

financial system from a less vulnerable state to a highly vulnerable state are a non-linear

function of the combination of network structures and bank-specific characteristics.34

Therefore, policies aiming at reducing systemic risk externalities related to

interconnectedness should focus on increasing the resilience of weak nodes in the system,

thereby curbing potential amplification effects due to cascade defaults, and on reshaping

the network structure in order to set a ceiling to potential losses.

As we show below, this requires a bank-specific calibration of prudential buffers

both targeting solvency and liquidity defaults so as to minimize amplification effects due

to contagion. Second-round default events are the key determinant of non-linear effects in

loss amplification, thereby becoming a natural candidate as an intermediate policy target

for macroprudential supervisors. In other words, the macroprudential regulator could aim

at reducing the role played by the network structure in terms of spreading contagion and

exposing the vulnerability of banks to shocks hitting the network. This paper proposes a

methodology to capture such amplification effects as reflected in the contagion and

vulnerability indexes (Figure 12) and their determinants (Tables 3 and 4). Moreover, as

we illustrate below, the CoMap methodology can be used to run counterfactual

simulations to study the effectiveness of different prudential actions in reducing

contagion potential in the network.

34 See also Battiston et al. (2009), Battiston and Caldarelli (2012) and Aikman et al. (2018).

50

For illustrative purposes, our counterfactual (macro) prudential simulations consist of

(i) increasing the buffer of HQLA assets (𝛾𝑖), (ii) increasing the pool of available for sale

assets (𝜃𝑖), and/or (iii) increasing the CET1 capital surplus (𝑘𝑖𝐷𝐹). The three distinct

policy measures aim at assessing the effectiveness of a single and equal increase in each

of the parameter in terms of reducing the number of banks experiencing a liquidity or

solvency default in Table 8.

Specifically, we consider a mix of the three policy measures, which is based on a

simple optimization problem for which we want to minimize the vulnerability level (VL)

- number of defaults experienced - so as to reduce the losses incurred due to second-

round effects (equation 21). This optimization problem is subject to an inequality

constraint which establishes that the sum of the parameters may not be larger than a

certain buffer (F) and each parameter should be at least equal to or higher than the

starting value �̅�, �̅� , 𝑘𝐷𝐹̅̅ ̅̅ ̅ derived by the bank-specific calibration.

min𝑉𝐿( 𝛾𝑖 , 𝜃𝑖 , 𝑘𝑖𝐷𝐹𝑖) (21)

𝑠. 𝑡. 𝜗1 𝑘𝑖𝐷𝐹 + 𝜗2 𝛾𝑖 + 𝜗3 𝜃𝑖 ≤ 𝐹 ; 𝑘𝑖

𝐷𝐹 ≥ 𝑘𝐷𝐹̅̅ ̅̅ ̅, 𝛾𝑖 ≥ �̅�, 𝜃𝑖 ≥ �̅�.

For simplicity, we set 𝜗1 = 𝜗2 = 𝜗3 = 1. 35 Figure 21 (panel a) presents the

vulnerability and contagion levels for the top-80 and top-40 banks with the highest

number of defaults experienced and induced, respectively. For the presentation of results,

the number of defaults is divided into four buckets capturing twenty (ten) banks each in a

descending order from most vulnerable (contagious) to less vulnerable.36

The benchmark

results are reported for comparative purposes.

The first exercise (LCR adj.) tests the effectiveness of an increase in the liquidity

buffer (𝛾𝑖) by 25 billion37

so as to increase the LCR ratio and better absorb the funding

shortfall induced by a default event. Since we are able to disentangle which banks suffer

35 For purely illustrative purposes, we assume (for simplicity) that the cost of increasing one buffer is

invariant across type of buffers. 36 It was preferred to display the top-80 instead of the top-50 in order to capture all banks defaulting. 37 The reasoning is that as initial policy rule we assume that banks experiencing a liquidity default would

increase the HQLA buffer up to the average of the sample (as depicted in Figure 8). This initial condition

increases the total amount of HQLA assets in the system by 25 billion. This policy rule is commonly used

in the literature, for instance see Gai et al. (2011). For the following exercises, this amount is shared across

the treated banks following a weighted average calculation based on total assets.

51

liquidity and solvency defaults, we threat with this policy experiment 8 banks

experiencing a default event. This experiment decreases the number of defaults by 16

units (red bars) out of 20 reported initially in the benchmark case (black bars). The

effectiveness of this treatment is relatively high, since the 4 remaining failures are all

solvency driven.

The second exercise (Pool adj.) resembles the first exercise and tests the

effectiveness in curbing liquidity driven defaults by increasing the pool of assets

available for sale by 25 billion. As captured by the blue bars, this policy rule is less

effective than an equal increase of the HQLA buffer, nevertheless it still reduces the

number of liquidity driven defaults by 11 units.

The third exercise (CET1 adj.) increases the average capital surplus by 25 billion.

Applying this policy measure reduces the number of defaults by only 6 units, although

the number of solvency driven defaults gets to 0.

The optimal policy mix set up by the minimization of the vulnerability level is able

to bring the total number of defaults among the top-80 most vulnerable banks to 0. Given

the set-up of the model, as described in equation (21), the most effective allocation is

divided between an increase in the liquidity buffer (𝛾𝑖) for those banks facing liquidity

driven defaults and an increase in the CET1 capital surplus for those banks vulnerable to

facing capital shortfalls. No fund is allocated to the pool of assets because each amount

used to absorb the funding shortfall will be discounted by a bank-specific discount rate,

leading to a less effective outcome than in the case of an equal increase in the buffer of

HQLA assets to which no discount rate is applied by assumption.38

Overall, the policy mix is able to bring the contagion level induced by the top-40

most systemic banks to zero. However, as shown in panel (b) of Figure 21, which reports

the vulnerability and contagion indexes, the scale of losses induced and experienced even

after the policy mix treatment still presents a fat-tailed distribution. This emphasizes that

decreasing the number of cascade defaults reduces contagion and vulnerability indexes

38 A different cost structure than 𝜗1 = 𝜗2 = 𝜗3 = 1 would affect the optimal policy mix.

52

by the contribution of amplification effects, which in our set-up are limited. In this

regard, first round effects dominate losses incurred due to cascade defaults.39

Figure 21: Comparative Statics of Policy Options

Panel (a)

Panel (b)

Note: Benchmark refers to results presented in section 4.1.

A final remark on the counterfactual policy simulations is that they focus solely on

the benefits related to curbing contagion risk in the system by modifying the network

structure and making banks more resilient. What is however not considered are the

potential costs of imposing more stringent requirements on banks (e.g. requiring higher

liquidity and or capital buffers). In reality, policy makers need to conduct a cost-benefit

analysis taking into account also the costs of doing so; for instance, in terms of lower

intermediation activity in the interbank market and other market segments where banks

interact.

39 This finding is in line with the literature; see e.g. Upper (2011) and Hüser (2015).

53

Conclusion

Our euro-centric systemic risk assessment based on the network of euro area banks’ large

exposures within the global banking system highlights that the degree of bank-specific

contagion and vulnerability depends on network specific tipping points affecting directly

the magnitude of amplification effects. This leads to the clear-cut conclusion that the

identification of such tipping points and their determinants is the essence of an effective

micro and macro prudential supervision. The current financial regulations seek to limit

each institution’s risk in isolation underestimating the systemic risk contribution to the

overall fragility.

In this paper, we argue that in isolation and with linear variations, bank-specific

characteristics seem to play a less relevant role than the network structure, whereas what

really determines a system shift comes from their non-linear interaction, for which both

are equally important. In a variety of tests, heterogeneity in the magnitude of bilateral

exposures and of bank-specific parameters is detected as a key driver of the total number

of defaults in the system. Unless systemic risk externalities are internalized by each bank

in the network, bank recapitalizations may be still convenient from a cost-return trade-off

of a global or European central planner. It follows that international cooperation is

essential to limit the ex-ante risk and reduce the ex-post system-wide losses striving for a

Pareto efficient outcome

Several extensions of our work should be explored in the future. The results are

network and model dependent based on an incomplete set of bilateral exposures.

Therefore, both dimensions need to be extended so as to include additional channels of

contagion and in turn improve the loss estimates of an extreme event. As a natural step

work should be done to incorporate i) euro area less significant institutions to complete

the euro area banking system, ii) financial corporations to model the complex interactions

within the financial system, and iii) exposures to real economy so as to capture feedback

loops. Moreover, we could complete the extra-euro area network by imputing missing

bilateral linkages by generating random networks consistent with partial information as in

Halaj and Kok (2013). Without changing the model assumptions, enlarging the dataset

dimension would lead to closer-to-reality second-round amplification effects. Next, the

modelling strategy may include a confidence channel so as to capture liquidity-hoarding

54

behaviors. This feature should bring funding risks to the forefront and determine a more

balanced contribution to loss estimation than the actual, which is mainly credit risk

driven. While we estimate fire-sales losses using a static balance-sheet approach, another

way to model them is dynamically by exploiting information on cross-holdings of assets

and derive the discount rate endogenously à la Cont and Schaanning (2017). Finally, we

should investigate the role of additional prudential requirements currently missing in our

framework such as a leverage ratio and a net stable funding ratio which in 2019 will

become binding.

Uncertainty surrounding the global financial network requires regulators to handle an

ever complex set of information so to be prepared in case such an adverse event takes

place. Nevertheless, networks are adaptive and so the policy mix needs to be. We have

provided a framework to capture few features of such complexity, and many more need

to be modelled to prune the fog of uncertainty, and take a decision when instability

suddenly arises.

55

References

Acharya V.V., R. Engle, and M. Richardson (2012), “Capital Shortfall: A New Approach

to Ranking and Regulating Systemic Risk”, American Economic Review: Papers &

Proceedings 102(3): 59-64

Acharya V.V., L.H. Pedersen, T. Philippon, and M. Richardson (2017), “Measuring

Systemic Risk”, The Review of Financial Studies 30(1):2-47.

Aikman D., Haldane A., Hinterschweiger M., and S. Kapadia (2018), “Rethinking

Financial Stability”, Staff Working Paper No. 712, Bank of England

Alter, A. and A. Beyer (2013), “The dynamics of spillover effects during the European

sovereign debt turmoil”, Journal of Banking and Finance 42: 134-153.

Anand K., B. Craig and G. von Peter (2014), “Filling in the Blanks: Network Structure

and Interbank Contagion”, Discussion Paper 02/2014, Deutsche Bundesbank

Bargigli L., F. Lillo, L. Infante and F. Pierobon (2015), “The Multiplex Structure of

Interbank Networks”, Quantitative Finance 15(4):673-691

Basu S., S. Das, G. Michailidis, A. Purnanandam (2017), A system-wide approach to

measure connectivity in the financial sector. Unpublished working paper.

Battiston S., J. Lorenz, and F. Schweitzer (2009), “Systemic Risk in a Unifying

Framework for Cascading Processes on Networks”, The European Physical Journal

B, 71(4): 441-460.

Battiston S. and S. Caldarelli (2012), “Systemic Risk in Financial Networks”, Journal of

Financial Management Markets and Institutions, 1(2): 129-254

Battiston, S., Delli Gatti, D., Gallegati, M., Greenwald, B., J.E. Stiglitz, (2012), “Default

cascades: when does risk diversification increase stability?” Journal of Financial

Stability, 8:138-149.

Billio M., M. Getmansky, A. W. Lo, and L. Pelizzon (2012), “Econometric Measures of

Connectedness and Systemic Risk in the Finance and Insurance Sectors”, Journal of

Financial Economics104 (3): 535–59

BIS - Bank for International Settlements (2014), Supervisory framework for measuring

and controlling large exposures. Standards, Basel Committee on Banking

Supervision

56

Brunnermeier M.K. and Y. Sannikov (2014), “A Macroeconomic Model with a Financial

Sector”, American Economic Review 104(2): 379-421.

Caballero R.J. and A. Simsek (2013), “Fire Sales in a Model of Complexity”, Journal of

Finance 68 (6): 2549–87

Caccioli F., M. Shrestha, C. Moore and J. D. Farmer (2014), “Stability Analysis of

Financial Contagion Due to Overlapping Portfolios”, Journal of Banking and

Finance 46: 233-45

Cifuentes, R., Ferrucci, G., H.S. Shin (2005), “Liquidity risk and contagion” Journal of

the European Economic Association, 3:556-566

Cont R. and E. Schaanning (2017), “Fire Sales, Indirect Contagion and Systemic Stress

Testing” Working paper Series No 2/2017, Norges Bank

Cont, R., Moussa, A., and E. Santos (2013), “Network Structure and Systemic Risk in

Banking Systems”, in J. Fouque & J. Langsam (Eds.), Handbook on Systemic Risk

(pp. 327-368). Cambridge: Cambridge University Press

Craig, B. and G. von Peter (2014), “Interbank Tiering and Money Center Banks”, Journal

of Financial Intermediation 23 (3): 322-47

Craig B., M. Koetter and U. Krüger (2014), “Interbank Lending and Distress:

Observables, Unobservables, and Network Structure”, Discussion Paper 18/2014,

Deutsche Bundesbank

CRR – Capital Requirements Regulation, regulation (EU) No 575/2013 of the European

Parliament and of the Council.

Demirer M., F. Diebold, L. Liu, K. Yilmaz (2017), “Estimating global bank network

Connectedness”, Journal of Applied Econometrics

Diebold, F., X., Yilmaz, K., 2009. Measuring Financial Asset Return and Volatility

Spillovers, with Application to Global Equity Markets. The Economic Journal

119(534), 158-171.

Diebold, F. X., Yilmaz, K., 2012. “Better to give than to receive: Predictive directional

measurement of volatility spillovers” International Journal of Forecasting (28): 57-

66

57

Diebold F.X. and K. Yilmaz (2014), “On the Network Topology of Variance

Decompositions: Measuring the Connectedness of Financial Firms”, Journal of

Econometrics 182 (1): 119-34

Degryse H. and G. Nguyen (2007) “Interbank Exposures: An Empirical Examination of

Contagion Risk in the Belgian Banking System”, International Journal of Central

Banking 3 (2): 123-71.

EBA/GL/2017/15 - European Banking Authority, Final Report: Guidelines on connected

clients under Article 4(1)(39) of Regulation (EU) No 575/2013.

Eisenberg L. and T.H. Noe (2001), “Systemic Risk in Financial Systems”, Management

Science 47(2): 236-49

Elsinger H., A. Lehar, and M. Summer (2006), “Risk Assessment for Banking Systems”,

Management Science 52 (9): 1301–14

Espinoza-Vega M.A. and J. Sole´ (2010), “Cross-Border Financial Surveillance: A

Network Perspective”, Working Paper Series 10/105, International Monetary Fund

Gai P. and S. Kapadia (2010), “Contagion in Financial Networks”, Proceedings of the

Royal Society Series A 466 (2120): 2401-23

Gai P., A. Haldane, and S. Kapadia (2011), “Complexity, Concentration and Contagion”,

Journal of Monetary Economics 58 (5): 453-70.

Garratt R.J., L. Mahadeva and K. Svirydzenka (2011), “Mapping systemic risk in the

international banking network” Working Paper No 413, Bank of England

Geithner, T. (2014), Stress Test: Reflections on Financial Crises, Crown Publishing

Group, Random House

Glasserman P. and H. P. Young (2016) “Contagion in Financial Network”, Journal of

Economic Literature 54(3): 779-831

Halaj G. and C. Kok (2013), “Assessing Interbank Contagion Using Simulated

Networks”, Computational Management Science 10 (2):157-86

Haldane A. G. and R.M. May (2011), “Systemic Risk in Banking Ecosystems.” Nature

469(7330): 351-55

Hüser A.C. (2015), “Too Interconnected to Fail: A Survey of the Interbank Networks

Literature”, Journal of Network Theory in Finance 1(3):1-50

58

Hüser A.C., G. Halaj, C. Kok, C. Perales, A. van der Kraaij (2017), “The Systemic

Implications of Bail-in: a Multi-layered Network Approach”, Working Paper Series

No 2010, European Central Bank

in’t Veld D. and I. van Lelyveld (2014), “Finding the Core: Network Structure in

Interbank Markets”, Journal of Banking and Finance 49: 27-40

Kok C., and M. Montagna (2016), “Multi-layered interbank model for assessing systemic

risk”, Working Paper Series No 1944, European Central Bank

Moratis G. and P. Sakellaris (2017), “Measuring the Systemic Importance of Banks”,

Working Paper Series No 240, Bank of Greece

Puhr C., R. Seliger and M. Sigmund (2012), “Contagiousness and Vulnerability in the

Austrian Interbank Market”, Financial Stability Report No 24, Oesterreichische

Nationalbank

Rogers L.C.G. and L.A.M. Veraart (2013), “Failure and Rescue in an Interbank

Network”, Management Science 59(4): 882–98.

Sheldon G., and M. Maurer (1998), “Interbank Lending and Systemic Risk: An Empirical

Analysis for Switzerland”, Swiss Journal of Economics and Statistics 134 (4): 685-

704

Upper C., 2011, “Simulation Methods to Assess the Danger of Contagion in Interbank

Markets”, Journal of Financial Stability 7 (3): 111-25

van Lelyveld I. and F. Liedorp, (2006) “Interbank Contagion in the Dutch Banking

Sector: A Sensitivity Analysis”, International Journal of Central Banking 2 (2): 99-

133.

59

Appendix

Data Infrastructure

Regarding the data reporting requirements, institutions should provide the name, Legal

Entity Identifier (LEI code), country, sector, and NACE classification of the

counterparty.40

This qualitative information is not always fulfilled, especially for

exposures to group of connected clients which cover approximately half of the data

sample. In addition, LEI codes are often missing and not every institution owns a LEI

code. This implies that identifying exposures vis-à-vis the very same counterparty across

countries can only take place through the counterparty’s name. However, the

counterparty’s name is often reported differently by different reporting institution and in

different languages according to the reporting country. For that reason, to exploit the

large exposures data and analyze the complex network of euro area banks’ large

exposures a reconciliation (or mapping) of counterparty names is necessary. Furthermore,

to complete the set of inputs for the modelling framework (section 3) is required to link

large exposures information with additional data sources.

In achieving this, we have developed an advanced algorithm defined as the Stata

Mapping Code (SMC), which aims at mapping counterparties’ names as well as filling

data reporting deficiencies regarding missing and misleading counterparty’s details (e.g.

LEI codes, country and sector)41

.

At our reference date (Q3 2017), before applying the SMC, there were almost 34.080

exposures of which 9.880 reported from euro area less significant institutions (LSIs). Out

of the 24.200 remaining exposures reported by euro area significant institutions (SIs),

15.363 are from institutions which are subsidiaries of the group, while the remaining

8.837 are from euro area consolidated group of SIs. These are the reporting institutions

40 The guidelines by the European Banking Authority (EBA) include common reporting templates and

guidance in relation to large exposures (LE) reporting within the COREP framework. There are three

templates included in the LE reporting framework that constitute the main source of data in establishing

bank network. These are: LE1 (C. 27.00) identification of the counterparty; LE2 (C. 28.00) exposures to

individual client and group of connected clients; LE3 (C. 29.00) detail of the exposures to individual clients

within groups of connected clients; LE4 (C. 30.00) and LE5 (C. 31.00): detail the information regarding the

maturity buckets to which the expected maturing amounts of the ten largest exposures to institutions as well

as the ten largest exposures to unregulated financial sector entities shall be allocated, respectively for table

C.28.00 and C.29.00. 41 The SMC follows a four-step approach. The step (1) aims at reconciling counterparties’ names using

fuzzy match commands based on a set of names’ similarities, existing LEI codes or combination of

counterparty-specific keywords and counterparty details such as country or sector of belonging. Step (2)

cross-merges the dataset by counterparties’ names and LEI codes in order to exploit existing

counterparties’ information so as to fill missing LEI codes, country and sectoral details. Step (3) enhances

further data quality by retrieving the missing counterparty information from the following sources:

gleif.org, Bloomberg or Bankscope. In the end, step (2) is repeated to complete the cycle. This cycle is

repeated up to the point no information is added to the system. Two cycles are sufficient to cover most of

the improvements.

60

on which we base this study; that is, the 107 euro area significant institutions at the

highest level of consolidation.42

In this regard, before applying the SMC, 5.833 individual counterparties were

identified, while after the algorithm mapping counterparty names was implemented the

number reduces by 1.353 units. Furthermore, the SMC by cross-merging information

already existing within the dataset and by adding missing information retrieved from

Bloomberg and Bankscope, is able to increase data availability on counterparties’ details

on average by 40%43

. The coverage of LEI codes increases from 30% to 69%, the

country dimension from 68% to 97%, sectors from 57% to 94%, and NACE codes from

34% to 81%.

Specifically, regarding the counterparty sector of interest - credit institutions - we identify

2.189 large exposures towards 498 unique counterparties, of which 77 are euro area

significant institutions (SIs), 155 non-euro area credit institutions, 238 euro area less-

significant institutions (LSIs), 21 are state development banks (SDBs), while 7 are

international organizations (IOs). In our study, we focus on large exposures between

reporting SIs and the first two counterparty groups, that is, SIs and non-euro area credit

institutions. Therefore, we drop from the sample of exposures LSIs, SDBs, and IOs. We

do this because for the modelling framework we need precise information about the

capital base and RWAs that are often not available for state development banks and

international organizations, as well as because exposures to SDBs and IOs are often

riskless. The inclusion of LSIs in the counterparty sample would be consistent only if we

include LSIs on the reporting side too. However, given the large number of LSI reporting

entities, we leave the LSI dimension for a future investigation.

42 For the sake of clarity, euro area SIs are 118 as cut-off date January 2018. However, since we work with

consolidated banking groups, large exposure amounts comprehend the exposures from its subsidiaries, i.e.,

from other SIs. For instance Unicredit Austria is a significant institution which belongs to Unicredit spa. In

this respect, the number of individual SIs decrease from 118 to 107 because of the consolidated approach.

See SSM’ list of supervised entities, cut-off date January 2018:

https://www.bankingsupervision.europa.eu/ecb/pub/pdf/ssm.list_of_supervised_entities_201802.en.pdf 43 We aim at having complete information about the counterparties for two main reasons: (1) LEI codes are

necessary to link the dataset with complementary sources of information about bank balance sheet data and

in turn calibrate the bank-specific model’s parameters; (2) country codes are necessary to disentangle the

geographical contribution to systemic risk.

61

Table A1: Top 30 Most Contagious and Vulnerable Banks – Special Cases

Note: Special case I sets ( 𝜆, 𝜌, 𝛾, 𝜃, 𝛿 ) equal to the average of the sample, while special case II sets ( 𝜆, 𝜌, 𝛿 ) equal to the average of the sample and 𝛾 = 0 and 𝜃 = ∞. CI EA refers to contagion index and

amounts represent capital losses to EA banks in percent of EA banking system’s total capital buffer. This

index is further decomposed into the respective contributions by credit (CI EA CR) and funding (CI EA

FU) shocks. VI refers to vulnerability index and amounts represent average capital losses across all

independent simulations in percent of a bank’s capital buffer. The vulnerability index is further

decomposed into the respective contributions by credit (VI CR) and funding (VI FU) shocks. The results in

the tables are ranked respectively according to CI EA and VI based on the benchmark model with bank-

specific parameters

Amp. Amp. Defaults Amp.

RANK Country Solv. Solv. Liq. CI CI CR CI FU Ratio Defaults Solv. Liq. CI CI CR CI FU Ratio ToT CI CI CR CI FU Ratio

1 EA 0 0 0 3.8 3.8 0.0 0.0 2 2 0 4.5 4.5 0.0 0.2 2 4.8 4.5 0.3 0.2

2 XEA 3 1 2 3.2 3.1 0.1 0.1 2 2 0 3.3 3.3 0.0 0.0 2 3.3 3.3 0.1 0.0

3 XEA 1 1 0 2.7 2.3 0.4 0.5 5 5 0 4.0 4.0 0.0 1.1 5 4.5 4.0 0.6 1.3

4 EA 2 2 0 2.4 2.4 0.0 0.0 6 6 0 4.7 4.7 0.0 1.1 6 5.1 4.7 0.5 1.1

5 EA 0 0 0 2.3 2.3 0.0 0.0 1 1 0 2.8 2.8 0.0 0.0 1 2.9 2.8 0.1 0.0

6 XEA 3 0 3 2.3 2.1 0.2 0.2 1 1 0 2.9 2.9 0.0 0.1 1 3.2 2.9 0.3 0.2

7 EA 1 0 1 2.2 2.1 0.1 0.0 0 0 0 2.6 2.6 0.0 0.0 6 5.2 4.2 1.0 1.2

8 XEA 0 0 0 1.9 1.9 0.0 0.0 0 0 0 1.9 1.9 0.0 0.0 0 2.0 1.9 0.0 0.0

9 XEA 0 0 0 1.9 1.9 0.0 0.0 0 0 0 2.0 2.0 0.0 0.0 0 2.0 2.0 0.0 0.0

10 EA 0 0 0 1.9 1.9 0.0 0.0 1 1 0 2.4 2.4 0.0 0.0 1 2.6 2.4 0.2 0.0

11 EA 1 1 0 1.8 1.8 0.0 0.0 1 1 0 2.0 2.0 0.0 0.0 1 2.2 2.0 0.2 0.0

12 XEA 1 0 1 1.4 1.2 0.2 0.1 0 0 0 1.3 1.3 0.0 0.0 0 1.3 1.3 0.1 0.0

13 XEA 2 0 2 1.4 1.3 0.1 0.3 0 0 0 1.5 1.5 0.0 0.0 0 1.5 1.5 0.0 0.0

14 EA 0 0 0 1.3 1.3 0.0 0.0 0 0 0 1.2 1.2 0.0 0.0 0 1.5 1.2 0.3 0.0

15 XEA 0 0 0 1.1 1.0 0.1 0.0 5 5 0 2.8 2.8 0.0 2.5 5 3.3 2.8 0.5 2.8

16 XEA 0 0 0 1.0 1.0 0.0 0.0 0 0 0 1.0 1.0 0.0 0.0 0 1.1 1.0 0.0 0.0

17 EA 0 0 0 1.0 1.0 0.0 0.0 1 1 0 1.2 1.2 0.0 0.0 1 1.3 1.2 0.1 0.0

18 EA 0 0 0 0.9 0.9 0.0 0.0 0 0 0 0.9 0.9 0.0 0.0 0 1.2 0.9 0.2 0.0

19 EA 1 0 1 0.8 0.8 0.0 0.1 0 0 0 0.7 0.7 0.0 0.0 1 0.8 0.7 0.1 0.0

20 XEA 0 0 0 0.8 0.8 0.0 0.0 0 0 0 0.8 0.8 0.0 0.0 0 0.8 0.8 0.0 0.0

21 XEA 0 0 0 0.8 0.8 0.0 0.0 0 0 0 0.8 0.8 0.0 0.0 0 0.9 0.8 0.1 0.0

22 EA 1 1 0 0.8 0.4 0.4 4.3 0 0 0 0.1 0.1 0.0 0.0 0 0.3 0.1 0.2 0.0

23 XEA 0 0 0 0.8 0.8 0.0 0.0 0 0 0 0.7 0.7 0.0 0.0 0 0.7 0.7 0.0 0.0

24 EA 0 0 0 0.8 0.8 0.0 0.0 0 0 0 1.2 1.2 0.0 0.0 0 1.4 1.2 0.1 0.0

25 XEA 0 0 0 0.7 0.7 0.0 0.0 0 0 0 0.8 0.8 0.0 0.0 0 0.8 0.8 0.0 0.0

26 EA 0 0 0 0.6 0.6 0.0 0.0 0 0 0 0.6 0.6 0.0 0.0 0 0.8 0.6 0.2 0.0

27 XEA 0 0 0 0.6 0.6 0.0 0.0 0 0 0 0.6 0.6 0.0 0.0 0 0.6 0.6 0.0 0.0

28 XEA 0 0 0 0.6 0.6 0.0 0.0 0 0 0 0.5 0.5 0.0 0.0 0 0.5 0.5 0.0 0.0

29 EA 0 0 0 0.5 0.5 0.0 0.0 0 0 0 0.7 0.7 0.0 0.0 0 0.8 0.7 0.1 0.0

30 EA 0 0 0 0.5 0.4 0.1 0.0 0 0 0 0.5 0.5 0.0 0.0 0 0.6 0.5 0.1 0.0c

0.53 0.20 0.33 1.43 1.37 0.05 0.19 0.83 0.83 0.00 1.70 1.70 0 0.17 1.07 1.93 1.75 0.18 0.23

Contagion

AVERAGE TOP 30

Special Case 1

Average Parameters

Special Case 2

Average Parameters ( θ = ∞ and γ = 0 )

Defaults Euro Area Defaults Euro Area Euro Area

Bank-Specific Perameters

Model-Parameter Heterogeneity

Amp. Amp. Defaults Amp.

RANK Country ToT Solv. Liq. VI VI CR VI FU Ratio ToT Solv. Liq. VI VI CR VI FU Ratio ToT VI VI CR VI FU Ratio

1 EA 2 2 0 1.8 0.7 1.2 0.0 4 4 0 2.8 2.8 0.0 0.1 5 3.3 2.9 0.4 0.1

2 EA 0 0 0 1.4 1.4 0.0 0.0 0 0 0 2.0 2.0 0.0 0.0 0 2.0 2.0 0.0 0.0

3 EA 0 0 0 1.1 1.1 0.0 0.0 0 0 0 1.6 1.6 0.0 0.6 1 1.7 1.7 0.0 0.7

4 EA 0 0 0 1.0 1.0 0.0 0.0 1 1 0 1.2 1.2 0.0 0.0 1 1.2 1.2 0.0 0.0

5 EA 0 0 0 0.9 0.9 0.0 0.0 0 0 0 0.8 0.8 0.0 0.0 0 0.8 0.8 0.0 0.0

6 XEA 0 0 0 0.8 0.8 0.0 0.4 0 0 0 1.0 1.0 0.0 0.6 0 1.2 1.1 0.2 0.8

7 EA 0 0 0 0.8 0.8 0.0 0.0 0 0 0 0.7 0.7 0.0 0.0 0 0.8 0.7 0.0 0.0

8 EA 0 0 0 0.7 0.7 0.0 0.0 0 0 0 0.8 0.8 0.0 0.0 0 0.8 0.8 0.0 0.0

9 EA 0 0 0 0.6 0.6 0.0 0.0 0 0 0 0.6 0.6 0.0 0.0 0 0.7 0.6 0.1 0.1

10 EA 2 0 2 0.6 0.5 0.1 0.0 0 0 0 0.4 0.4 0.0 0.0 0 0.6 0.4 0.2 0.0

11 EA 0 0 0 0.6 0.6 0.0 0.2 6 6 0 6.4 6.4 0.0 0.9 7 7.0 6.4 0.6 0.9

12 EA 0 0 0 0.5 0.5 0.0 0.0 0 0 0 0.6 0.6 0.0 0.4 0 0.6 0.6 0.0 0.4

13 EA 0 0 0 0.5 0.5 0.0 0.0 0 0 0 0.4 0.4 0.0 0.0 0 0.5 0.4 0.0 0.0

14 EA 0 0 0 0.5 0.5 0.0 0.0 0 0 0 0.5 0.5 0.0 0.1 0 0.6 0.5 0.0 0.1

15 EA 0 0 0 0.5 0.5 0.0 0.1 0 0 0 0.5 0.5 0.0 0.2 0 0.6 0.5 0.0 0.2

16 EA 0 0 0 0.4 0.4 0.0 0.1 5 5 0 3.8 3.8 0.0 1.0 6 4.3 4.3 0.0 1.2

17 EA 0 0 0 0.4 0.4 0.0 0.0 0 0 0 0.4 0.4 0.0 0.0 0 0.5 0.4 0.1 0.1

18 EA 3 3 0 0.4 0.4 0.0 0.0 3 3 0 0.4 0.4 0.0 0.0 3 0.4 0.4 0.0 0.0

19 EA 0 0 0 0.4 0.4 0.0 0.0 0 0 0 0.5 0.5 0.0 0.0 0 0.5 0.5 0.0 0.0

20 EA 0 0 0 0.4 0.4 0.0 0.0 0 0 0 0.7 0.7 0.0 0.0 0 0.8 0.7 0.0 0.0

21 EA 0 0 0 0.4 0.4 0.0 0.0 0 0 0 0.3 0.3 0.0 0.0 0 0.4 0.3 0.1 0.1

22 EA 0 0 0 0.4 0.4 0.0 0.0 0 0 0 0.3 0.3 0.0 0.0 0 0.3 0.3 0.0 0.0

23 EA 0 0 0 0.4 0.4 0.0 0.0 0 0 0 0.4 0.4 0.0 0.1 0 0.4 0.4 0.1 0.2

24 EA 0 0 0 0.4 0.4 0.0 0.0 0 0 0 0.3 0.3 0.0 0.0 0 0.3 0.3 0.0 0.0

25 EA 0 0 0 0.4 0.4 0.0 0.0 0 0 0 0.6 0.6 0.0 0.0 0 0.7 0.6 0.1 0.0

26 EA 0 0 0 0.4 0.4 0.0 0.0 7 7 0 1.7 1.7 0.0 0.7 8 2.0 1.9 0.1 0.9

27 EA 0 0 0 0.3 0.3 0.0 0.0 0 0 0 0.3 0.3 0.0 0.0 0 0.4 0.4 0.0 0.0

28 EA 0 0 0 0.3 0.3 0.0 0.0 0 0 0 0.3 0.3 0.0 0.1 0 0.4 0.3 0.0 0.1

29 EA 0 0 0 0.3 0.3 0.0 0.0 1 1 0 0.3 0.3 0.0 0.0 1 0.4 0.3 0.0 0.0

30 EA 0 0 0 0.3 0.3 0.0 0.0 0 0 0 0.6 0.6 0.0 0.6 0 0.7 0.6 0.1 0.6

0.23 0.17 0.07 0.60 0.56 0.04 0.02 0.90 0.90 0.00 1.05 1.05 0.00 0.18 1.07 1.16 1.09 0.08 0.22

Average Parameters ( θ = ∞ and γ = 0 )

Special Case IISpecial Case I

Defaults Global Global

Average ParametersVulnerability

AVERAGE TOP 30

Defaults Global

Model-Parameter Heterogeneity

Bank-Specific Perameters

62

Figure 16: Distress vs Default CET1 Thresholds

Panel (a)

Panel (b)

Note: Panel (a) is ordered according to CET1 distress threshold. For confidentiality reasons, the chart

shows statistics as average among three banks.

Figure 17: CET1 Distress Threshold vs Total Capital Distress Threshold

Panel (a)

Panel (b)

Note: Panel (a) is ordered according to CET1 distress threshold. For confidentiality reasons, the chart

shows statistics as average among three banks.

63

Figure 18: Sensitivity to Funding Parameters

Panel (a): Funding Shortfall

Panel (b): Net Liquidity Position

Panel (c): Pool of assets

Note: Benchmark refers to results presented in section 4.1.

64

Figure 19: Network Structure Sensitivity Based on Gross Exposures

Note: Network Structured is based on the statistics presented in table 1 regarding gross exposures amounts

before taking into account exemptions. Benchmark refers to results presented in section 4.1.

Figure 20: Multi-Factor Sensitivity

Note: The liquidity scenario has been modelled as (∆𝜌𝑖 = 40%; ∆𝑁𝐿𝑃𝑖 = −20%; 𝜃𝑖 = −20%), the

capital scenario includes a 20% decrease in the capital surplus, while in the third scenario (network)

exposure amounts increase by 30% up to which (€ 1.120 billion).

65

Figure 21: SRISK Ranking Comparison

Note: Regarding the SRISK ranking, bank number 1 is the one with the highest SRISK estimate, i.e. the

one in the top-right corner of the SRISK Index. MES based SRISK estimates for European banks have been

retrieved from vlab.stern.nyu.edu. For confidentiality reasons, the chart shows statistics as average among

three banks.